增加nanodet、更新各平台模型

model/AX620E/yolov8n_npu1.axmodel → CPP/ax_nanodetplus_qrcode_batch

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f0d52b2b3692be7e00e9ce710ed3155e47499a0f580749f66ffe9cf32230bfb7
-size 3548343
+oid sha256:b736080ed5b459822c615e815c7f5914a1b21ea1d20644780ffcf5d74ac82813
+size 6424712

README.md

@@ -9,7 +9,7 @@ This version of QRCode detetion model has been converted to run on the Axera NPU
 
 This model has been optimized with the following LoRA: 
 
-Compatible with Pulsar2 version: 4.2
+Compatible with Pulsar2 version: 5.1
 
 ## Convert tools links:
 
@@ -19,7 +19,7 @@ For those who are interested in model conversion, you can try to export axmodel
 
 - [Pulsar2 Link, How to Convert ONNX to axmodel](https://pulsar2-docs.readthedocs.io/en/latest/pulsar2/introduction.html) 
 
-- [The repo of AXera Platform](https://github.com/AXERA-TECH/ax-samples),which you can compile the c++ demo
+- [The repo of AXera Platform](https://github.com/AXERA-TECH/ax-samples),which you can learn how to compile the C++ demo
 
 ## Support Platform
 
@@ -30,19 +30,34 @@ For those who are interested in model conversion, you can try to export axmodel
   - [爱芯派2](https://axera-pi-2-docs-cn.readthedocs.io/zh-cn/latest/index.html)
   - [Module-LLM](https://docs.m5stack.com/zh_CN/module/Module-LLM)
   - [LLM630 Compute Kit](https://docs.m5stack.com/zh_CN/core/LLM630%20Compute%20Kit)
+- AX637
 
 |Chips|model|cost|
 |--|--|--|
-||yolov5n|0.73 ms|
-||yolov8n|1.31 ms|
-||yolov9t|1.89 ms|
-|AX650|yolov10n|1.44 ms|
-||yolo11n|1.39 ms|
-||yolo12n|2.49 ms|
-||DEIMv2_femto(u16)|1.79 ms|
-
-|AX630C|yolov5n|2.57 ms|
-|AX630C|yolov8n|5.89 ms|
+||yolov5n|1.73 ms|
+||yolov8n|3.64 ms|
+||yolov9t|4.75 ms|
+|AX650|yolov10n|3.67 ms|
+||yolo11n|3.42 ms|
+||yolo12n|6.87 ms|
+||NanodetPlus|2.16 ms|
+||DEIMv2_femto(u16)|3.76 ms|
+|||
+||yolov5n|5.79 ms|
+||yolov8n|9.26 ms|
+||yolov9t|11.6 ms|
+|AX630C|yolov10n|9.71 ms|
+||yolo11n|9.65 ms|
+||yolo12n|20.24 ms|
+||NanodetPlus|5.93 ms|
+|||
+||yolov5n|2.11 ms|
+||yolov8n|4.04 ms|
+||yolov9t|4.91 ms|
+|AX637|yolov10n|4.05 ms|
+||yolo11n|3.84 ms|
+||yolo12n|6.40 ms|
+||NanodetPlus|2.38 ms|
 
 ## How to use
 
@@ -50,46 +65,59 @@ Download all files from this repository to the device
 
 ```
 
-root@ax650:~/QRCode_det# tree
 .
-|-- CPP
-|   |-- ax_deimv2_qrcode_batch
-|   |-- ax_yolov5_qrcode_batch
-|   `-- ax_yolov8_qrcode_batch
-|-- README.md
-|-- images
-|   |-- qrcode_01.jpg
-|   |-- qrcode_02.jpg
-|   |-- qrcode_03.jpg
-|   ....
-|   `-- qrcode_55.jpg
-|-- model
-|   |-- AX620E
-|   |   |-- yolov5n_npu1.axmodel
-|   |   |-- yolov5n_npu2.axmodel
-|   |   |-- yolov8n_npu1.axmodel
-|   |   `-- yolov8n_npu2.axmodel
-|   |-- AX637
-|   |   |-- deimv2_hgnetv2_femto_coco_npu1.axmodel
-|   |   |-- yolov5n_npu1.axmodel
-|   |   `-- yolov8n_npu1.axmodel
-|   |-- AX650
-|   |   |-- deimv2_hgnetv2_femto_coco_npu3.axmodel
-|   |   |-- yolov5n_npu3.axmodel
-|   |   `-- yolov8n_npu3.axmodel
-|   `-- CPP
-|       |-- deimv2_hgnetv2_femto_coco_cpp_npu3.axmodel
-|       |-- yolov5n_cpp_npu3.axmodel
-|       `-- yolov8n_cpp_npu3.axmodel
-|-- python
-|   |-- QRCode_axmodel_infer_DEIMv2.py
-|   |-- QRCode_axmodel_infer_v5.py
-|   |-- QRCode_axmodel_infer_v8.py
-|   |-- QRCode_onnx_infer_DEIMv2.py
-|   |-- QRCode_onnx_infer_v5.py
-|   |-- QRCode_onnx_infer_v8.py
-|   `-- requirements.txt
-`-- result.png
+├── config.json
+├── CPP
+│   ├── ax_deimv2_qrcode_batch
+│   ├── ax_nanodetplus_qrcode_batch
+│   ├── ax_yolov5_qrcode_batch
+│   └── ax_yolov8_qrcode_batch
+├── cpp_result.png
+├── images
+│   ├── qrcode_01.jpg
+│   ├── qrcode_02.jpg
+│   ├── qrcode_03.jpg
+|   ├── ...
+│   └── qrcode_55.jpg
+├── model
+│   ├── AX620E
+│   │   ├── nanodet-plus-m_630_npu1.axmodel
+│   │   ├── yolo11n_630_npu1.axmodel
+│   │   ├── yolo12n_630_npu1.axmodel
+│   │   ├── yolov10n_630_npu1.axmodel
+│   │   ├── yolov5n_630_npu1.axmodel
+│   │   ├── yolov8n_630_npu1.axmodel
+│   │   └── yolov9t_630_npu1.axmodel
+│   ├── AX637
+│   │   ├── nanodet-plus-m_637_npu1.axmodel
+│   │   ├── yolo11n_637_npu1.axmodel
+│   │   ├── yolo12n_637_npu1.axmodel
+│   │   ├── yolov10n_637_npu1.axmodel
+│   │   ├── yolov5n_637_npu1.axmodel
+│   │   ├── yolov8n_637_npu1.axmodel
+│   │   └── yolov9t_637_npu1.axmodel
+│   └── AX650
+│       ├── deimv2_femto_650_npu1_u16.axmodel
+│       ├── nanodet-plus-m_650_npu1.axmodel
+│       ├── yolo11n_650_npu1.axmodel
+│       ├── yolo12n_650_npu1.axmodel
+│       ├── yolov10n_650_npu1.axmodel
+│       ├── yolov5n_650_npu1.axmodel
+│       ├── yolov8n_650_npu1.axmodel
+│       └── yolov9t_650_npu1.axmodel
+├── py_result.png
+├── python
+│   ├── QRCode_axmodel_infer_DEIMv2.py
+│   ├── QRCode_axmodel_infer_Nanodet.py
+│   ├── QRCode_axmodel_infer_v5.py
+│   ├── QRCode_axmodel_infer_v8.py
+│   ├── QRCode_onnx_infer_DEIMv2.py
+│   ├── QRCode_onnx_infer_Nanodet.py
+│   ├── QRCode_onnx_infer_v5.py
+│   ├── QRCode_onnx_infer_v8.py
+│   └���─ requirements.txt
+└── README.md
+
 
 ```
 
@@ -135,7 +163,7 @@ Output:
 
 ##### C++
 ```
-./ax_xxx_qrcode_batch -m xxx_cpp_npu3.axmodel -i images/ -o res/
+./ax_xxx_qrcode_batch -m xxx_npu1.axmodel -i images/
 ```
 
 Output:

model/AX620E/{yolov8n_npu2.axmodel → nanodet-plus-m_630_npu1.axmodel}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e18042a394d440cc6fc73a7b35871a189081961f3a5cb72301e76edc9799d3a3
-size 3208843
+oid sha256:5480ca6e968f13fc3146cea39d3d5043e251fa9d7c4971a3ec813daefbd67a5b
+size 1899764

model/AX620E/{yolov5n_npu1.axmodel → yolo11n_630_npu1.axmodel}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cb93d21015ca639308bdd42cc3275c5776cfdd5d67c822b08d1d10843db767a9
-size 2089800
+oid sha256:0af92b5e73b0c419421c58be8c942b48e199358ec021714e11830b764a5d8d43
+size 3290722

model/AX620E/{yolov5n_npu2.axmodel → yolo12n_630_npu1.axmodel}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a73396b9bc57a097554cf5d9c54f745b2ea4c0f24a3644f0317d13cd415bb4e2
-size 1854620
+oid sha256:f57663e403e10ec989405d8e9cd93bea32c38ce1e67e6fcc9f1bcfd486bc9782
+size 4468598

model/AX620E/yolov10n_630_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5749f1768cf1f3f01f345975453ac924d81b8005b558ff0cd0fe5dd4b6f96bd8
+size 2952122

model/AX620E/yolov5n_630_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:178447ef748f9e89668094b588b42d64e28a3dccee50247548cbaceac2edf253
+size 2070769

model/AX620E/yolov8n_630_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6449c6f85e949e2621906807b3a36e7c02aa1cb03748def2d813dbdff285ead7
+size 3548856

model/AX620E/yolov9t_630_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d861ef093708e85e2b3c8bcf7963343c3583d33fba47cc3fece00fd8e6a09571
+size 2979512

model/AX637/deimv2_hgnetv2_femto_coco_npu1.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1a15e46fcca5f762a7076aefc8eab68bb9db0d072dd42270b215968633d9a009
-size 1568889

model/AX637/nanodet-plus-m_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e1cafbbae5cc1ea1491170518dcb7daffd9ff04b3f95514f0a1fd1f56f78351
+size 3584897

model/AX637/yolo11n_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c020f4d48a8196292422336c47ba125bf1752ca645b42052e480c4cd65a73fb
+size 2874030

model/AX637/yolo12n_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c62fde3b1ff6d2c01a4580a89b797e19be52ebceb9d5053f3e45ffcd6191d36
+size 4164182

model/AX637/yolov10n_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2bda66e4b6b69bdc8ed88b55d0f5e41faca9c88dfa51a055c1cefa234cc5f4
+size 3005816

model/AX637/yolov5n_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8ba900b4a1a6e37f3ef9890305a6d35d01d06b7a8c3822ea872291ebfd5fa2b
+size 1865855

model/AX637/yolov5n_npu1.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:13f44b48e782769a57aba01e07d6724f60a507dbdb0ecf35e44f3838ac58798e
-size 1865062

model/AX637/yolov8n_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c9dd054118ca8b52041accd0c5a5147e6344f91a68b26c35e3f98d9460db558
+size 3178420

model/AX637/yolov8n_npu1.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:80a03632e1c762094f7bbc6115674270b5c23db8fcc8dcdaf65443988b6e0292
-size 3178331

model/AX637/yolov9t_637_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7bbdcce782f6f56fc6a0dd3aa9206a1fdfbc5cafc10c6a180f48a9aea878b5f
+size 2771528

model/AX650/deimv2_femto_650_npu1_u16.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd252aa527a23eb1794d11af264d4650dfe9248eac84f87a5bfb555a3f23a7aa
+size 1753703

model/AX650/deimv2_hgnetv2_femto_coco_npu3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:1bd80d8338290a4e88c26bf994604559c094838227286382959966c497db738e
-size 2204278

model/AX650/nanodet-plus-m_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:765665a81790917431660f6c5c5a3f57ea6e4360951c863c8fac5e5c7e6f655d
+size 2120369

model/AX650/yolo11n_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb71cb0f04a65c489425d1b6c83a28a7510cde99062d81147106e5b798805647
+size 2995284

model/AX650/yolo12n_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6a743c1f7d9476cf284b4cb4f7c0731fd366a81d8e4d34fc9a334cf4a157603
+size 3392100

model/AX650/yolov10n_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c0a84073c9488b0943f8e2c0c479583e1c89bd8589e31646767418e788a5533
+size 2735472

model/AX650/yolov5n_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:358edd177ed91eda32047cee72283d11201e4d1996a44e5631008abb224e015e
+size 1925943

model/AX650/yolov5n_npu3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:34ea4a3e74bd9dde388086cb60cd51f9d53c7734bf52f2e21de9a6d71de0b194
-size 2003219

model/AX650/yolov8n_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c10cde32e29ecc33395e2cfeab4645c83c9a45afcb0e946f02993a30bcf63a0
+size 3247992

model/AX650/yolov8n_npu3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d05be5bad116511e552849e16793b8b3ea19238a8fa1dc00935a18fb58207e81
-size 3488708

model/AX650/yolov9t_650_npu1.axmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fcf218afc4ae555daa137570e3ff16088ffd9b3de5fc5537690ba93091ca5b8
+size 2706764

model/CPP/deimv2_hgnetv2_femto_coco_cpp_npu3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a7559ef2f367d70e54b46cdbdf2fa39e139387064799cc73f00b613a30b3133d
-size 2197074

model/CPP/yolov5n_cpp_npu3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:2a59402253a3e4f14c6ff93d939f6bdb889b86858dcd729c6f2d82efe96f6da1
-size 2001946

model/CPP/yolov8n_cpp_npu3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:44eef3ec1e8d5d7cb8925dcfe4b25dd1826445baa18b0cf4db4deb5e87f7412e
-size 3487467

python/QRCode_axmodel_infer_DEIMv2.py

@@ -227,18 +227,19 @@ def process_image(sess, im_pil, post_processor, size=640, model_size='s'):
     resized_im_pil, ratio, pad_w, pad_h = resize_with_aspect_ratio(im_pil, size)
     orig_size = torch.tensor([[resized_im_pil.size[1], resized_im_pil.size[0]]])
 
-    transforms = T.Compose([
-            T.ToTensor(),
-            T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) 
-                if model_size not in ['atto', 'femto', 'pico', 'n'] 
-                else T.Lambda(lambda x: x)
-        ])
-
-    im_data = transforms(resized_im_pil).unsqueeze(0)
-
+    # transforms = T.Compose([
+    #         T.ToTensor(),
+    #         T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) 
+    #             if model_size not in ['atto', 'femto', 'pico', 'n'] 
+    #             else T.Lambda(lambda x: x)
+    #     ])
+
+    # im_data = transforms(resized_im_pil).unsqueeze(0)
+    im_data = np.array(resized_im_pil)
+    im_data = np.expand_dims(im_data, axis=0).astype(np.uint8)
     output = sess.run(
         output_names=None,
-        input_feed={'images': im_data.numpy()}
+        input_feed={'images': im_data}
     )
 
     output = {"pred_logits": torch.from_numpy(output[0]), "pred_boxes": torch.from_numpy(output[1])}
@@ -262,10 +263,10 @@ class QRCodeDecoder:
         for idx, region in enumerate(regions):
             x1, y1, x2, y2 = region
             # 外扩缓解检测截断，视检测情况而定
-            # x1-=15
-            # y1-=15
-            # x2+=15
-            # y2+=15
+            x1-=15
+            y1-=15
+            x2+=15
+            y2+=15
             # 裁剪图像
             cropped = image[y1:y2, x1:x2]
             if cropped.size > 0:
@@ -309,12 +310,12 @@ class QRCodeDecoder:
 if __name__ == '__main__':
 
     #load the ONNX model
-    sess = axe.InferenceSession('deimv2_hgnetv2_femto_coco_npu3.axmodel')
+    sess = axe.InferenceSession('deimv2_femto_650_npu1_u16.axmodel')
     size = sess.get_inputs()[0].shape[2]
 
     #QRCode decoder
     decoder = QRCodeDecoder()
-    img_path = './images'
+    img_path = './qrcode_test'
     det_path='./DEIMv2_det_res'
     crop_path='./DEIMv2_crop_res'
 

python/QRCode_axmodel_infer_Nanodet.py

@@ -0,0 +1,715 @@
+import os
+import glob
+import time
+import cv2
+import numpy as np
+import pyzbar.pyzbar as pyzbar
+import axengine as axe
+import math
+
+names = ["QRCode"]
+
+def sigmoid(x):
+    return 1 / (1 + np.exp(-x))
+def model_load(model):
+    session = axe.InferenceSession(model)
+    input_name = session.get_inputs()[0].name
+    output_names = [ x.name for x in session.get_outputs()]
+    return session, output_names
+
+def data_process_cv2(frame, input_shape):
+    im0 = cv2.imread(frame)
+    img = cv2.resize(im0, input_shape, interpolation=cv2.INTER_AREA)
+    org_data = img.copy()
+    img = np.ascontiguousarray(img)
+    img = np.expand_dims(img, 0)
+    return img, im0, org_data
+    
+
+def multiclass_nms(
+    multi_bboxes, multi_scores, score_thr, nms_cfg, max_num=-1, score_factors=None
+):
+    num_classes = multi_scores.shape[1] - 1  # exclude background
+
+    # Reshape bboxes
+    if multi_bboxes.shape[1] > 4:
+        # (N, 4*C) -> (N, C, 4)
+        bboxes = multi_bboxes.reshape(multi_scores.shape[0], -1, 4)
+    else:
+        # (N, 4) -> (N, 1, 4) -> (N, C, 4) via repeat
+        bboxes = np.tile(multi_bboxes[:, None, :], (1, num_classes, 1))
+
+    scores = multi_scores[:, :-1].copy()  # (N, C)
+
+    # Apply score factors if provided
+    if score_factors is not None:
+        scores = scores * score_factors[:, None]
+
+    # Filter by score threshold
+    valid_mask = scores > score_thr  # (N, C)
+
+    # Get indices where valid
+    valid_indices = np.where(valid_mask)
+    if len(valid_indices[0]) == 0:
+        # No valid boxes
+        return np.zeros((0, 5), dtype=np.float32), np.zeros((0,), dtype=np.int64)
+
+    # Extract valid bboxes, scores, labels
+    bbox_indices, class_indices = valid_indices
+    bboxes_valid = bboxes[bbox_indices, class_indices]  # (K, 4)
+    scores_valid = scores[valid_indices]                # (K,)
+    labels_valid = class_indices.astype(np.int64)       # (K,)
+
+    # Concatenate bboxes and scores for NMS input: (K, 5)
+    dets_input = np.concatenate([bboxes_valid, scores_valid[:, None]], axis=1)  # (K, 5)
+
+    # Perform NMS (you need a NumPy NMS implementation)
+    keep = nms_numpy(dets_input, iou_threshold=nms_cfg.get('iou_threshold', 0.5))
+
+    dets = dets_input[keep]
+    labels = labels_valid[keep]
+
+    if max_num > 0 and len(keep) > max_num:
+        dets = dets[:max_num]
+        labels = labels[:max_num]
+
+    return dets, labels
+def nms_numpy(dets, iou_threshold=0.5):
+    if dets.size == 0:
+        return []
+
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    scores = dets[:, 4]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]  # descending order
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+
+        iou = inter / (areas[i] + areas[order[1:]] - inter)
+        inds = np.where(iou <= iou_threshold)[0]
+        order = order[inds + 1]
+
+    return keep
+def batched_nms(boxes, scores, idxs, nms_cfg, class_agnostic=False):
+    nms_cfg_ = nms_cfg.copy()
+    class_agnostic = nms_cfg_.pop("class_agnostic", class_agnostic)
+
+    if class_agnostic:
+        boxes_for_nms = boxes
+    else:
+        max_coordinate = boxes.max()
+        # offsets = idxs * (max_coordinate + 1)
+        offsets = idxs.astype(boxes.dtype) * (max_coordinate + 1)
+        boxes_for_nms = boxes + offsets[:, None]
+
+    nms_type = nms_cfg_.pop("type", "nms")  # unused in numpy version
+    split_thr = nms_cfg_.pop("split_thr", 10000)
+
+    if len(boxes_for_nms) < split_thr:
+        # Call your NumPy NMS function (e.g., nms_numpy)
+        keep = nms_numpy(boxes_for_nms, scores, **nms_cfg_)
+        keep = np.array(keep, dtype=np.int64)
+        boxes = boxes[keep]
+        scores = scores[keep]
+    else:
+        # Large case: process per class/group
+        total_mask = np.zeros(scores.shape, dtype=bool)
+        unique_ids = np.unique(idxs)
+
+        for id_val in unique_ids:
+            mask = (idxs == id_val)
+            mask_indices = np.where(mask)[0]  # indices where condition is True
+
+            if len(mask_indices) == 0:
+                continue
+
+            keep_in_group = nms_numpy(
+                boxes_for_nms[mask_indices], 
+                scores[mask_indices], 
+                **nms_cfg_
+            )
+            keep_in_group = np.array(keep_in_group, dtype=np.int64)
+            selected_global_indices = mask_indices[keep_in_group]
+            total_mask[selected_global_indices] = True
+
+        keep = np.where(total_mask)[0]
+        # Sort by scores descending
+        sorted_indices = np.argsort(-scores[keep])  # negative for descending
+        keep = keep[sorted_indices]
+        boxes = boxes[keep]
+        scores = scores[keep]
+
+    # Concatenate boxes and scores -> (K, 5)
+    dets = np.concatenate([boxes, scores[:, None]], axis=-1)
+    return dets, keep
+
+def scale_boxes_no_letter(img1_shape, boxes, img0_shape):
+    gain = (img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])
+
+    boxes[..., [0, 2]] /= gain[1]
+    boxes[..., [1, 3]] /= gain[0]
+    clip_boxes(boxes, img0_shape)
+    return boxes
+
+def clip_boxes(boxes, shape):
+    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])
+    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])
+
+_COLORS = (
+    np.array(
+        [
+            0.000,
+            0.447,
+            0.741,
+            0.850,
+            0.325,
+            0.098,
+            0.929,
+            0.694,
+            0.125,
+            0.494,
+            0.184,
+            0.556,
+            0.466,
+            0.674,
+            0.188,
+            0.301,
+            0.745,
+            0.933,
+            0.635,
+            0.078,
+            0.184,
+            0.300,
+            0.300,
+            0.300,
+            0.600,
+            0.600,
+            0.600,
+            1.000,
+            0.000,
+            0.000,
+            1.000,
+            0.500,
+            0.000,
+            0.749,
+            0.749,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            1.000,
+            0.667,
+            0.000,
+            1.000,
+            0.333,
+            0.333,
+            0.000,
+            0.333,
+            0.667,
+            0.000,
+            0.333,
+            1.000,
+            0.000,
+            0.667,
+            0.333,
+            0.000,
+            0.667,
+            0.667,
+            0.000,
+            0.667,
+            1.000,
+            0.000,
+            1.000,
+            0.333,
+            0.000,
+            1.000,
+            0.667,
+            0.000,
+            1.000,
+            1.000,
+            0.000,
+            0.000,
+            0.333,
+            0.500,
+            0.000,
+            0.667,
+            0.500,
+            0.000,
+            1.000,
+            0.500,
+            0.333,
+            0.000,
+            0.500,
+            0.333,
+            0.333,
+            0.500,
+            0.333,
+            0.667,
+            0.500,
+            0.333,
+            1.000,
+            0.500,
+            0.667,
+            0.000,
+            0.500,
+            0.667,
+            0.333,
+            0.500,
+            0.667,
+            0.667,
+            0.500,
+            0.667,
+            1.000,
+            0.500,
+            1.000,
+            0.000,
+            0.500,
+            1.000,
+            0.333,
+            0.500,
+            1.000,
+            0.667,
+            0.500,
+            1.000,
+            1.000,
+            0.500,
+            0.000,
+            0.333,
+            1.000,
+            0.000,
+            0.667,
+            1.000,
+            0.000,
+            1.000,
+            1.000,
+            0.333,
+            0.000,
+            1.000,
+            0.333,
+            0.333,
+            1.000,
+            0.333,
+            0.667,
+            1.000,
+            0.333,
+            1.000,
+            1.000,
+            0.667,
+            0.000,
+            1.000,
+            0.667,
+            0.333,
+            1.000,
+            0.667,
+            0.667,
+            1.000,
+            0.667,
+            1.000,
+            1.000,
+            1.000,
+            0.000,
+            1.000,
+            1.000,
+            0.333,
+            1.000,
+            1.000,
+            0.667,
+            1.000,
+            0.333,
+            0.000,
+            0.000,
+            0.500,
+            0.000,
+            0.000,
+            0.667,
+            0.000,
+            0.000,
+            0.833,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            0.167,
+            0.000,
+            0.000,
+            0.333,
+            0.000,
+            0.000,
+            0.500,
+            0.000,
+            0.000,
+            0.667,
+            0.000,
+            0.000,
+            0.833,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            0.167,
+            0.000,
+            0.000,
+            0.333,
+            0.000,
+            0.000,
+            0.500,
+            0.000,
+            0.000,
+            0.667,
+            0.000,
+            0.000,
+            0.833,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            0.143,
+            0.143,
+            0.143,
+            0.286,
+            0.286,
+            0.286,
+            0.429,
+            0.429,
+            0.429,
+            0.571,
+            0.571,
+            0.571,
+            0.714,
+            0.714,
+            0.714,
+            0.857,
+            0.857,
+            0.857,
+            0.000,
+            0.447,
+            0.741,
+            0.314,
+            0.717,
+            0.741,
+            0.50,
+            0.5,
+            0,
+        ]
+    )
+    .astype(np.float32)
+    .reshape(-1, 3)
+)
+
+def distance2bbox(points, distance, max_shape=None):
+    x1 = points[..., 0] - distance[..., 0]
+    y1 = points[..., 1] - distance[..., 1]
+    x2 = points[..., 0] + distance[..., 2]
+    y2 = points[..., 1] + distance[..., 3]
+    if max_shape is not None:
+        x1 = np.clip(x1, a_min=0, a_max=max_shape[1])
+        y1 = np.clip(y1, a_min=0, a_max=max_shape[0])
+        x2 = np.clip(x2, a_min=0, a_max=max_shape[1])
+        y2 = np.clip(y2, a_min=0, a_max=max_shape[0])
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+def integral_numpy(x, reg_max=16):
+    """
+    NumPy equivalent of the Integral layer in NanoDet.
+    
+    Computes: sum(softmax(logits) * [0, 1, ..., reg_max]) for each of the 4 directions.
+    
+    Args:
+        x (np.ndarray): Input array of shape (..., 4 * (reg_max + 1))
+        reg_max (int): Maximum value of discrete set. Default: 16.
+    
+    Returns:
+        np.ndarray: Integral result of shape (..., 4)
+    """
+    # Save original leading shape (e.g., (N,) or (N, H, W))
+    leading_shape = x.shape[:-1]  # everything except last dim
+    total_channels = x.shape[-1]
+    
+    assert total_channels == 4 * (reg_max + 1), \
+        f"Last dimension must be 4*(reg_max+1)={4*(reg_max+1)}, but got {total_channels}"
+    
+    # Reshape to (..., 4, reg_max + 1)
+    x = x.reshape(*leading_shape, 4, reg_max + 1)
+    
+    # Apply softmax along the last axis (dim=-1)
+    # For numerical stability: subtract max
+    x_max = np.max(x, axis=-1, keepdims=True)
+    exp_x = np.exp(x - x_max)
+    softmax_x = exp_x / np.sum(exp_x, axis=-1, keepdims=True)  # (..., 4, reg_max+1)
+    
+    # Project vector: [0, 1, 2, ..., reg_max]
+    project = np.arange(reg_max + 1, dtype=x.dtype)  # shape (reg_max+1,)
+    
+    # Compute weighted sum: sum(softmax_x * project) over last dimension
+    # Broadcasting: (..., 4, reg_max+1) * (reg_max+1,) -> (..., 4, reg_max+1)
+    integral_result = np.sum(softmax_x * project, axis=-1)  # (..., 4)
+    
+    return integral_result
+
+def overlay_bbox_cv(img, dets, class_names, score_thresh):
+    all_box = []
+    for label in dets:
+        for bbox in dets[label]:
+            score = bbox[-1]
+            if score > score_thresh:
+                x0, y0, x1, y1 = [int(i) for i in bbox[:4]]
+                all_box.append([label, x0, y0, x1, y1, score])
+    all_box.sort(key=lambda v: v[5])
+    # for box in all_box:
+    #     label, x0, y0, x1, y1, score = box
+    #     # color = self.cmap(i)[:3]
+    #     color = (_COLORS[label] * 255).astype(np.uint8).tolist()
+    #     text = "{}:{:.1f}%".format(class_names[label], score * 100)
+    #     txt_color = (0, 0, 0) if np.mean(_COLORS[label]) > 0.5 else (255, 255, 255)
+    #     font = cv2.FONT_HERSHEY_SIMPLEX
+    #     txt_size = cv2.getTextSize(text, font, 0.5, 2)[0]
+        # cv2.rectangle(img, (x0, y0), (x1, y1), color, 2)
+
+        # cv2.rectangle(
+        #     img,
+        #     (x0, y0 - txt_size[1] - 1),
+        #     (x0 + txt_size[0] + txt_size[1], y0 - 1),
+        #     color,
+        #     -1,
+        # )
+        # cv2.putText(img, text, (x0, y0 - 1), font, 0.5, txt_color, thickness=1)
+    return img, all_box
+    
+class NanoDetONNXInfer:
+    def __init__(self, model_path, imgsz=[416, 416]):
+        self.model_path = model_path
+        self.session, self.output_names = model_load(self.model_path)
+        self.imgsz = imgsz
+        self.reg_max = 7
+        self.reg_max1= self.reg_max + 1
+        self.distribution_project = np.arange(self.reg_max + 1)
+        self.nc = len(names)
+        self.no = self.nc + self.reg_max1 * 4
+        self.stride = [8, 16, 32, 64]
+
+    def get_bboxes(self, cls_preds, reg_preds):
+        """Decode the outputs to bboxes.
+        Args:
+            cls_preds (Tensor): Shape (num_imgs, num_points, num_classes).
+            reg_preds (Tensor): Shape (num_imgs, num_points, 4 * (regmax + 1)).
+            img_metas (dict): Dict of image info.
+
+        Returns:
+            results_list (list[tuple]): List of detection bboxes and labels.
+        """
+        b = cls_preds.shape[0]
+
+        featmap_sizes = [
+            (math.ceil(self.imgsz[0] / stride), math.ceil(self.imgsz[1]) / stride)
+            for stride in self.stride
+        ]
+
+        # get grid cells of one image
+        mlvl_center_priors = [
+            self.get_single_level_center_priors(
+                b,
+                featmap_sizes[i],
+                stride,
+                dtype=np.float32,
+            )
+            for i, stride in enumerate(self.stride)
+        ]
+
+        center_priors = np.concatenate(mlvl_center_priors, axis=1)
+        integral_result = integral_numpy(reg_preds, reg_max=self.reg_max)  # (N, 4)
+        scale = center_priors[..., 2][..., None]  # shape (N, 1) or (N, H, W, 1)
+        dis_preds = integral_result * scale
+        bboxes = distance2bbox(center_priors[..., :2], dis_preds, max_shape=self.imgsz)
+        scores = 1.0 / (1.0 + np.exp(-cls_preds))  # sigmoid
+        result_list = []
+        for i in range(b):
+            # add a dummy background class at the end of all labels
+            # same with mmdetection2.0
+            score, bbox = scores[i], bboxes[i]
+            padding = np.zeros((score.shape[0], 1), dtype=score.dtype)
+            score = np.concatenate([score, padding], axis=1)
+            results = multiclass_nms(
+                bbox,
+                score,
+                score_thr=0.05,
+                nms_cfg=dict(type="nms", iou_threshold=0.6),
+                max_num=100,
+            )
+            result_list.append(results)
+        return result_list
+    def get_single_level_center_priors(self,batch_size, featmap_size, stride, dtype):
+        h, w = featmap_size
+        x_range = (np.arange(w, dtype=dtype)) * stride
+        y_range = (np.arange(h, dtype=dtype)) * stride
+        y, x = np.meshgrid(y_range, x_range, indexing='ij')
+        y = y.flatten()
+        x = x.flatten()
+        strides = np.full((x.shape[0],), stride, dtype=dtype)
+        priors = np.stack([x, y, strides, strides], axis=-1)
+        return np.tile(priors[None, :, :], (batch_size, 1, 1))
+
+    def detect_objects(self, image, save_path):
+        outputs=[]
+        im, im0, org_data = data_process_cv2(image, self.imgsz)
+        img_name = os.path.basename(image).split('.')[0]
+        infer_start_time = time.time()
+        x = self.session.run(None, {self.session.get_inputs()[0].name: im})
+        infer_end_time = time.time()
+        print(f"infer time: {infer_end_time - infer_start_time:.4f}s")
+        x = [np.transpose(x[i],(0,3,1,2)) for i in range(4)]    #to nchw
+        for i in range(len(x)):
+            reg_pred = x[i][:, :self.reg_max1 * 4,:,:]
+            cls_pred = x[i][:, self.reg_max1 * 4:,:,:]
+            out = np.concatenate([cls_pred, reg_pred], axis=1)
+            outputs.append(out.reshape(out.shape[0], out.shape[1], -1))
+        preds = np.concatenate(outputs, axis=2).transpose(0, 2, 1)
+
+        cls_scores = preds[:, :, :self.nc]
+        bbox_preds = preds[:, :, self.nc:]
+        pred = self.get_bboxes(cls_scores, bbox_preds)[0]
+        res = self.post_process(pred, org_data, im0, save_path, img_name)
+        result_img, bbox_res = overlay_bbox_cv(im0, res, names, score_thresh=0.35)
+        return bbox_res, result_img
+    def post_process(self, result, im, im0, save_path, img_name):
+        det_result = {}
+        det_bboxes, det_labels = result
+        det_bboxes[:, :4] = scale_boxes_no_letter(im.shape[:2], det_bboxes[:, :4], im0.shape).round()
+        classes = det_labels
+        for i in range(self.nc):
+            inds = classes == i
+            det_result[i] = np.concatenate(
+                [
+                    det_bboxes[inds, :4].astype(np.float32),
+                    det_bboxes[inds, 4:5].astype(np.float32),
+                ],
+                axis=1,
+            ).tolist()
+
+        return det_result
+    
+class QRCodeDecoder:
+    def crop_qr_regions(self, image, regions):
+        """
+        根据检测到的边界框裁剪二维码区域
+        """
+        cropped_images = []
+        for idx, region in enumerate(regions):
+            label, x1, y1, x2, y2, score = region
+            # 外扩15个像素缓解因检测截断造成无法识别的情况，视检测情况而定
+            x1-=15
+            y1-=15
+            x2+=15
+            y2+=15
+            # 裁剪图像
+            cropped = image[y1:y2, x1:x2]
+            if cropped.size > 0:
+                cropped_images.append({
+                    'image': cropped,
+                    'bbox': region,
+                })
+        return cropped_images
+
+    def decode_qrcode_pyzbar(self, cropped_image):
+        """
+        使用pyzbar解码二维码
+        """
+        try:
+            # 转换为灰度图像
+            if len(cropped_image.shape) == 3:
+                gray = cv2.cvtColor(cropped_image, cv2.COLOR_BGR2GRAY)
+            else:
+                gray = cropped_image
+            # 使用pyzbar解码
+            decoded_objects = pyzbar.decode(gray)
+            results = []
+            for obj in decoded_objects:
+                try:
+                    data = obj.data.decode('utf-8')
+                    results.append({
+                        'data': data,
+                        'type': obj.type,
+                        'points': obj.polygon
+                    })
+                except:
+                    continue
+            
+            return results
+        except Exception as e:
+            print(f"decode error: {e}")
+            return []
+
+if __name__ == '__main__':
+    import time
+
+    detector = NanoDetONNXInfer(model_path='./nanodet-plus-m_416_QR.axmodel',imgsz=[416,416])
+    decoder = QRCodeDecoder()
+    img_path = './qrcode_test'
+    det_path='./det_res'
+    crop_path='./crop_res'
+    os.makedirs(det_path, exist_ok=True)
+    os.makedirs(crop_path, exist_ok=True)
+    imgs = glob.glob(f"{img_path}/*.jpg")
+    totoal = len(imgs)
+    success = 0
+    fail = 0
+    start_time = time.time()
+    for idx,img in enumerate(imgs):
+        pic_name=os.path.basename(img).split('.')[0]
+        loop_start_time = time.time()
+        det_result, res_img = detector.detect_objects(img,det_path)
+        # cv2.imwrite(os.path.join(det_path, pic_name+'.jpg'), res_img)
+        # print('det_result:',det_result)
+        # Crop deteted QRCode & decode QRCode by pyzbar
+        cropped_images = decoder.crop_qr_regions(res_img, det_result)
+        # for i,cropped in enumerate(cropped_images):
+        #     cv2.imwrite(os.path.join(crop_path, f'{pic_name}_crop_{i}.jpg'), cropped['image'])
+
+        all_decoded_results = []
+        for i, cropped_data in enumerate(cropped_images):
+            decoded_results = decoder.decode_qrcode_pyzbar(cropped_data['image'])
+            all_decoded_results.extend(decoded_results)
+            
+            # for result in decoded_results:
+            #     print(f"decode result: {result['data']} (type: {result['type']})")
+        if all_decoded_results:
+            success += 1
+            print(f"{pic_name} 识别成功！")
+        else:
+            fail += 1
+            print(f"{pic_name} 识别失败！")
+        loop_end_time = time.time()
+        print(f"图片 {img} 处理耗时: {loop_end_time - loop_start_time:.4f} 秒")
+
+    end_time = time.time()  # 记录总结束时间
+    total_time = end_time - start_time  # 记录总耗时
+
+    print(f"总共测试图片数量: {totoal}")
+    print(f"识别成功数量: {success}")
+    print(f"识别失败数量: {fail}")
+    print(f"识别成功率: {success/totoal*100:.2f}%")
+    print(f"整体处理耗时: {total_time:.4f} 秒")
+    print(f"平均每张图片处理耗时: {total_time/totoal:.4f} 秒")

python/QRCode_axmodel_infer_v5.py

@@ -280,9 +280,10 @@ class Yolov5QRcodeDetector:
 
     def preprocess_image(self, img, img_size=(640, 640)):
         img, _, _ = letterbox(img, img_size, auto=False, stride=32)
-        img = np.ascontiguousarray(img[:, :, ::-1].transpose(2, 0, 1))
+        # img = np.ascontiguousarray(img[:, :, ::-1].transpose(2, 0, 1))
+        img = np.ascontiguousarray(img).astype(np.uint8)
         # img = np.asarray(img, dtype=np.float32)
-        img = np.asarray(img, dtype=np.uint8)
+        # img = np.asarray(img, dtype=np.uint8)
         img = np.expand_dims(img, 0)
         # img /= 255.0
         return img
@@ -300,6 +301,7 @@ class Yolov5QRcodeDetector:
         grid = torch.stack((xv, yv), 2).expand(shape) - 0.5  # add grid offset, i.e. y = 2.0 * x - 0.5
         anchor_grid = (self.anchors[i] * self.stride[i]).view((1, na, 1, 1, 2)).expand(shape)
         return grid, anchor_grid
+
     def postprocess(self, preds, img_shape, im0):
         z = []  # inference output
         for i,pred in enumerate(preds):
@@ -322,7 +324,8 @@ class Yolov5QRcodeDetector:
 
             if len(det):
                 # Rescale boxes from img_size to im0 size
-                scale_coords(img_shape[2:], det[:, :4], im0.shape, kpt_label=False)
+                # scale_coords(img_shape[2:], det[:, :4], im0.shape, kpt_label=False)
+                scale_coords(img_shape[1:3], det[:, :4], im0.shape, kpt_label=False)
 
                 # Print results
                 for c in det[:, 5].unique():
@@ -397,12 +400,12 @@ class QRCodeDecoder:
 if __name__ == '__main__':
     import time
 
-    model = './yolov5n_npu3.axmodel'
+    model = './yolov5n_650_npu1.axmodel'
     input_size = [640,640]
     detector = Yolov5QRcodeDetector(model)
     # Crop deteted QRCode & decode QRCode by pyzbar
     decoder = QRCodeDecoder()
-    pic_path = './images/'
+    pic_path = './qrcode_test/'
     det_path='./v5_det_res'
     crop_path='./v5_crop_res'
     os.makedirs(det_path, exist_ok=True)

python/QRCode_axmodel_infer_v8.py

@@ -49,7 +49,8 @@ def data_process_cv2(frame, input_shape):
     im0 = cv2.imread(frame)
     img = letterbox(im0, input_shape, auto=False, stride=32)[0]
     org_data = img.copy()
-    img = np.ascontiguousarray(img[:, :, ::-1].transpose(2, 0, 1))
+    # img = np.ascontiguousarray(img[:, :, ::-1].transpose(2, 0, 1))
+    img = np.ascontiguousarray(img[:, :, ::-1])
     img = np.asarray(img, dtype=np.uint8)
     img = np.expand_dims(img, 0)
     # img /= 255.0
@@ -506,9 +507,9 @@ class QRCodeDecoder:
 if __name__ == '__main__':
     import time
 
-    detector = YOLOV8Detector(model_path='./yolov8n_npu3.axmodel',imgsz=[640,640])
+    detector = YOLOV8Detector(model_path='./yolov8n_650_npu1.axmodel',imgsz=[640,640])
     decoder = QRCodeDecoder()
-    img_path = './images'
+    img_path = './qrcode_test'
     det_path='./v8_det_res'
     crop_path='./v8_crop_res'
     os.makedirs(det_path, exist_ok=True)

python/QRCode_onnx_infer_Nanodet.py

@@ -0,0 +1,718 @@
+import os
+import glob
+import time
+import cv2
+import numpy as np
+import pyzbar.pyzbar as pyzbar
+import onnxruntime as ort
+import math
+
+names = ["QRCode"]
+
+def sigmoid(x):
+    return 1 / (1 + np.exp(-x))
+def model_load(model):
+    providers = ['CPUExecutionProvider']
+    session = ort.InferenceSession(model, providers=providers)
+    input_name = session.get_inputs()[0].name
+    output_names = [ x.name for x in session.get_outputs()]
+    return session, output_names
+
+def data_process_cv2(frame, input_shape):
+    mean = np.array([103.53, 116.28, 123.675], dtype=np.float32).reshape(1, 1, 3)
+    std = np.array([57.375, 57.12, 58.395], dtype=np.float32).reshape(1, 1, 3)
+    im0 = cv2.imread(frame)
+    img = cv2.resize(im0, input_shape, interpolation=cv2.INTER_AREA).astype(np.float32)
+    org_data = img.copy()
+    img = (img - mean) / std
+    img = np.ascontiguousarray(img.transpose(2, 0, 1))
+    img = np.expand_dims(img, 0)
+    return img, im0, org_data
+
+def multiclass_nms(
+    multi_bboxes, multi_scores, score_thr, nms_cfg, max_num=-1, score_factors=None
+):
+    num_classes = multi_scores.shape[1] - 1  # exclude background
+
+    # Reshape bboxes
+    if multi_bboxes.shape[1] > 4:
+        # (N, 4*C) -> (N, C, 4)
+        bboxes = multi_bboxes.reshape(multi_scores.shape[0], -1, 4)
+    else:
+        # (N, 4) -> (N, 1, 4) -> (N, C, 4) via repeat
+        bboxes = np.tile(multi_bboxes[:, None, :], (1, num_classes, 1))
+
+    scores = multi_scores[:, :-1].copy()  # (N, C)
+
+    # Apply score factors if provided
+    if score_factors is not None:
+        scores = scores * score_factors[:, None]
+
+    # Filter by score threshold
+    valid_mask = scores > score_thr  # (N, C)
+
+    # Get indices where valid
+    valid_indices = np.where(valid_mask)
+    if len(valid_indices[0]) == 0:
+        # No valid boxes
+        return np.zeros((0, 5), dtype=np.float32), np.zeros((0,), dtype=np.int64)
+
+    # Extract valid bboxes, scores, labels
+    bbox_indices, class_indices = valid_indices
+    bboxes_valid = bboxes[bbox_indices, class_indices]  # (K, 4)
+    scores_valid = scores[valid_indices]                # (K,)
+    labels_valid = class_indices.astype(np.int64)       # (K,)
+
+    # Concatenate bboxes and scores for NMS input: (K, 5)
+    dets_input = np.concatenate([bboxes_valid, scores_valid[:, None]], axis=1)  # (K, 5)
+
+    # Perform NMS (you need a NumPy NMS implementation)
+    keep = nms_numpy(dets_input, iou_threshold=nms_cfg.get('iou_threshold', 0.5))
+
+    dets = dets_input[keep]
+    labels = labels_valid[keep]
+
+    if max_num > 0 and len(keep) > max_num:
+        dets = dets[:max_num]
+        labels = labels[:max_num]
+
+    return dets, labels
+def nms_numpy(dets, iou_threshold=0.5):
+    if dets.size == 0:
+        return []
+
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    scores = dets[:, 4]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]  # descending order
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+
+        iou = inter / (areas[i] + areas[order[1:]] - inter)
+        inds = np.where(iou <= iou_threshold)[0]
+        order = order[inds + 1]
+
+    return keep
+def batched_nms(boxes, scores, idxs, nms_cfg, class_agnostic=False):
+    nms_cfg_ = nms_cfg.copy()
+    class_agnostic = nms_cfg_.pop("class_agnostic", class_agnostic)
+
+    if class_agnostic:
+        boxes_for_nms = boxes
+    else:
+        max_coordinate = boxes.max()
+        # offsets = idxs * (max_coordinate + 1)
+        offsets = idxs.astype(boxes.dtype) * (max_coordinate + 1)
+        boxes_for_nms = boxes + offsets[:, None]
+
+    nms_type = nms_cfg_.pop("type", "nms")  # unused in numpy version
+    split_thr = nms_cfg_.pop("split_thr", 10000)
+
+    if len(boxes_for_nms) < split_thr:
+        # Call your NumPy NMS function (e.g., nms_numpy)
+        keep = nms_numpy(boxes_for_nms, scores, **nms_cfg_)
+        keep = np.array(keep, dtype=np.int64)
+        boxes = boxes[keep]
+        scores = scores[keep]
+    else:
+        # Large case: process per class/group
+        total_mask = np.zeros(scores.shape, dtype=bool)
+        unique_ids = np.unique(idxs)
+
+        for id_val in unique_ids:
+            mask = (idxs == id_val)
+            mask_indices = np.where(mask)[0]  # indices where condition is True
+
+            if len(mask_indices) == 0:
+                continue
+
+            keep_in_group = nms_numpy(
+                boxes_for_nms[mask_indices], 
+                scores[mask_indices], 
+                **nms_cfg_
+            )
+            keep_in_group = np.array(keep_in_group, dtype=np.int64)
+            selected_global_indices = mask_indices[keep_in_group]
+            total_mask[selected_global_indices] = True
+
+        keep = np.where(total_mask)[0]
+        # Sort by scores descending
+        sorted_indices = np.argsort(-scores[keep])  # negative for descending
+        keep = keep[sorted_indices]
+        boxes = boxes[keep]
+        scores = scores[keep]
+
+    # Concatenate boxes and scores -> (K, 5)
+    dets = np.concatenate([boxes, scores[:, None]], axis=-1)
+    return dets, keep
+
+def scale_boxes_no_letter(img1_shape, boxes, img0_shape):
+    gain = (img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])
+
+    boxes[..., [0, 2]] /= gain[1]
+    boxes[..., [1, 3]] /= gain[0]
+    clip_boxes(boxes, img0_shape)
+    return boxes
+
+def clip_boxes(boxes, shape):
+    boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])
+    boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])
+
+_COLORS = (
+    np.array(
+        [
+            0.000,
+            0.447,
+            0.741,
+            0.850,
+            0.325,
+            0.098,
+            0.929,
+            0.694,
+            0.125,
+            0.494,
+            0.184,
+            0.556,
+            0.466,
+            0.674,
+            0.188,
+            0.301,
+            0.745,
+            0.933,
+            0.635,
+            0.078,
+            0.184,
+            0.300,
+            0.300,
+            0.300,
+            0.600,
+            0.600,
+            0.600,
+            1.000,
+            0.000,
+            0.000,
+            1.000,
+            0.500,
+            0.000,
+            0.749,
+            0.749,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            1.000,
+            0.667,
+            0.000,
+            1.000,
+            0.333,
+            0.333,
+            0.000,
+            0.333,
+            0.667,
+            0.000,
+            0.333,
+            1.000,
+            0.000,
+            0.667,
+            0.333,
+            0.000,
+            0.667,
+            0.667,
+            0.000,
+            0.667,
+            1.000,
+            0.000,
+            1.000,
+            0.333,
+            0.000,
+            1.000,
+            0.667,
+            0.000,
+            1.000,
+            1.000,
+            0.000,
+            0.000,
+            0.333,
+            0.500,
+            0.000,
+            0.667,
+            0.500,
+            0.000,
+            1.000,
+            0.500,
+            0.333,
+            0.000,
+            0.500,
+            0.333,
+            0.333,
+            0.500,
+            0.333,
+            0.667,
+            0.500,
+            0.333,
+            1.000,
+            0.500,
+            0.667,
+            0.000,
+            0.500,
+            0.667,
+            0.333,
+            0.500,
+            0.667,
+            0.667,
+            0.500,
+            0.667,
+            1.000,
+            0.500,
+            1.000,
+            0.000,
+            0.500,
+            1.000,
+            0.333,
+            0.500,
+            1.000,
+            0.667,
+            0.500,
+            1.000,
+            1.000,
+            0.500,
+            0.000,
+            0.333,
+            1.000,
+            0.000,
+            0.667,
+            1.000,
+            0.000,
+            1.000,
+            1.000,
+            0.333,
+            0.000,
+            1.000,
+            0.333,
+            0.333,
+            1.000,
+            0.333,
+            0.667,
+            1.000,
+            0.333,
+            1.000,
+            1.000,
+            0.667,
+            0.000,
+            1.000,
+            0.667,
+            0.333,
+            1.000,
+            0.667,
+            0.667,
+            1.000,
+            0.667,
+            1.000,
+            1.000,
+            1.000,
+            0.000,
+            1.000,
+            1.000,
+            0.333,
+            1.000,
+            1.000,
+            0.667,
+            1.000,
+            0.333,
+            0.000,
+            0.000,
+            0.500,
+            0.000,
+            0.000,
+            0.667,
+            0.000,
+            0.000,
+            0.833,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            0.167,
+            0.000,
+            0.000,
+            0.333,
+            0.000,
+            0.000,
+            0.500,
+            0.000,
+            0.000,
+            0.667,
+            0.000,
+            0.000,
+            0.833,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            0.167,
+            0.000,
+            0.000,
+            0.333,
+            0.000,
+            0.000,
+            0.500,
+            0.000,
+            0.000,
+            0.667,
+            0.000,
+            0.000,
+            0.833,
+            0.000,
+            0.000,
+            1.000,
+            0.000,
+            0.000,
+            0.000,
+            0.143,
+            0.143,
+            0.143,
+            0.286,
+            0.286,
+            0.286,
+            0.429,
+            0.429,
+            0.429,
+            0.571,
+            0.571,
+            0.571,
+            0.714,
+            0.714,
+            0.714,
+            0.857,
+            0.857,
+            0.857,
+            0.000,
+            0.447,
+            0.741,
+            0.314,
+            0.717,
+            0.741,
+            0.50,
+            0.5,
+            0,
+        ]
+    )
+    .astype(np.float32)
+    .reshape(-1, 3)
+)
+
+def distance2bbox(points, distance, max_shape=None):
+    x1 = points[..., 0] - distance[..., 0]
+    y1 = points[..., 1] - distance[..., 1]
+    x2 = points[..., 0] + distance[..., 2]
+    y2 = points[..., 1] + distance[..., 3]
+    if max_shape is not None:
+        x1 = np.clip(x1, a_min=0, a_max=max_shape[1])
+        y1 = np.clip(y1, a_min=0, a_max=max_shape[0])
+        x2 = np.clip(x2, a_min=0, a_max=max_shape[1])
+        y2 = np.clip(y2, a_min=0, a_max=max_shape[0])
+    return np.stack([x1, y1, x2, y2], axis=-1)
+
+def integral_numpy(x, reg_max=16):
+    """
+    NumPy equivalent of the Integral layer in NanoDet.
+    
+    Computes: sum(softmax(logits) * [0, 1, ..., reg_max]) for each of the 4 directions.
+    
+    Args:
+        x (np.ndarray): Input array of shape (..., 4 * (reg_max + 1))
+        reg_max (int): Maximum value of discrete set. Default: 16.
+    
+    Returns:
+        np.ndarray: Integral result of shape (..., 4)
+    """
+    # Save original leading shape (e.g., (N,) or (N, H, W))
+    leading_shape = x.shape[:-1]  # everything except last dim
+    total_channels = x.shape[-1]
+    
+    assert total_channels == 4 * (reg_max + 1), \
+        f"Last dimension must be 4*(reg_max+1)={4*(reg_max+1)}, but got {total_channels}"
+    
+    # Reshape to (..., 4, reg_max + 1)
+    x = x.reshape(*leading_shape, 4, reg_max + 1)
+    
+    # Apply softmax along the last axis (dim=-1)
+    # For numerical stability: subtract max
+    x_max = np.max(x, axis=-1, keepdims=True)
+    exp_x = np.exp(x - x_max)
+    softmax_x = exp_x / np.sum(exp_x, axis=-1, keepdims=True)  # (..., 4, reg_max+1)
+    
+    # Project vector: [0, 1, 2, ..., reg_max]
+    project = np.arange(reg_max + 1, dtype=x.dtype)  # shape (reg_max+1,)
+    
+    # Compute weighted sum: sum(softmax_x * project) over last dimension
+    # Broadcasting: (..., 4, reg_max+1) * (reg_max+1,) -> (..., 4, reg_max+1)
+    integral_result = np.sum(softmax_x * project, axis=-1)  # (..., 4)
+    
+    return integral_result
+
+def overlay_bbox_cv(img, dets, class_names, score_thresh):
+    all_box = []
+    for label in dets:
+        for bbox in dets[label]:
+            score = bbox[-1]
+            if score > score_thresh:
+                x0, y0, x1, y1 = [int(i) for i in bbox[:4]]
+                all_box.append([label, x0, y0, x1, y1, score])
+    all_box.sort(key=lambda v: v[5])
+    # for box in all_box:
+    #     label, x0, y0, x1, y1, score = box
+    #     # color = self.cmap(i)[:3]
+    #     color = (_COLORS[label] * 255).astype(np.uint8).tolist()
+    #     text = "{}:{:.1f}%".format(class_names[label], score * 100)
+    #     txt_color = (0, 0, 0) if np.mean(_COLORS[label]) > 0.5 else (255, 255, 255)
+    #     font = cv2.FONT_HERSHEY_SIMPLEX
+    #     txt_size = cv2.getTextSize(text, font, 0.5, 2)[0]
+        # cv2.rectangle(img, (x0, y0), (x1, y1), color, 2)
+
+        # cv2.rectangle(
+        #     img,
+        #     (x0, y0 - txt_size[1] - 1),
+        #     (x0 + txt_size[0] + txt_size[1], y0 - 1),
+        #     color,
+        #     -1,
+        # )
+        # cv2.putText(img, text, (x0, y0 - 1), font, 0.5, txt_color, thickness=1)
+    return img, all_box
+    
+class NanoDetONNXInfer:
+    def __init__(self, model_path, imgsz=[416, 416]):
+        self.model_path = model_path
+        self.session, self.output_names = model_load(self.model_path)
+        self.imgsz = imgsz
+        self.reg_max = 7
+        self.reg_max1= self.reg_max + 1
+        self.distribution_project = np.arange(self.reg_max + 1)
+        self.nc = len(names)
+        self.no = self.nc + self.reg_max1 * 4
+        self.stride = [8, 16, 32, 64]
+
+    def get_bboxes(self, cls_preds, reg_preds):
+        """Decode the outputs to bboxes.
+        Args:
+            cls_preds (Tensor): Shape (num_imgs, num_points, num_classes).
+            reg_preds (Tensor): Shape (num_imgs, num_points, 4 * (regmax + 1)).
+            img_metas (dict): Dict of image info.
+
+        Returns:
+            results_list (list[tuple]): List of detection bboxes and labels.
+        """
+        b = cls_preds.shape[0]
+
+        featmap_sizes = [
+            (math.ceil(self.imgsz[0] / stride), math.ceil(self.imgsz[1]) / stride)
+            for stride in self.stride
+        ]
+
+        # get grid cells of one image
+        mlvl_center_priors = [
+            self.get_single_level_center_priors(
+                b,
+                featmap_sizes[i],
+                stride,
+                dtype=np.float32,
+            )
+            for i, stride in enumerate(self.stride)
+        ]
+
+        center_priors = np.concatenate(mlvl_center_priors, axis=1)
+        integral_result = integral_numpy(reg_preds, reg_max=self.reg_max)  # (N, 4)
+        scale = center_priors[..., 2][..., None]  # shape (N, 1) or (N, H, W, 1)
+        dis_preds = integral_result * scale
+        bboxes = distance2bbox(center_priors[..., :2], dis_preds, max_shape=self.imgsz)
+        scores = 1.0 / (1.0 + np.exp(-cls_preds))  # sigmoid
+        result_list = []
+        for i in range(b):
+            # add a dummy background class at the end of all labels
+            # same with mmdetection2.0
+            score, bbox = scores[i], bboxes[i]
+            padding = np.zeros((score.shape[0], 1), dtype=score.dtype)
+            score = np.concatenate([score, padding], axis=1)
+            results = multiclass_nms(
+                bbox,
+                score,
+                score_thr=0.05,
+                nms_cfg=dict(type="nms", iou_threshold=0.6),
+                max_num=100,
+            )
+            result_list.append(results)
+        return result_list
+    def get_single_level_center_priors(self,batch_size, featmap_size, stride, dtype):
+        h, w = featmap_size
+        x_range = (np.arange(w, dtype=dtype)) * stride
+        y_range = (np.arange(h, dtype=dtype)) * stride
+        y, x = np.meshgrid(y_range, x_range, indexing='ij')
+        y = y.flatten()
+        x = x.flatten()
+        strides = np.full((x.shape[0],), stride, dtype=dtype)
+        priors = np.stack([x, y, strides, strides], axis=-1)
+        return np.tile(priors[None, :, :], (batch_size, 1, 1))
+
+    def detect_objects(self, image, save_path):
+        outputs=[]
+        im, im0, org_data = data_process_cv2(image, self.imgsz)
+        img_name = os.path.basename(image).split('.')[0]
+        infer_start_time = time.time()
+        x = self.session.run(None, {self.session.get_inputs()[0].name: im})
+        infer_end_time = time.time()
+        print(f"infer time: {infer_end_time - infer_start_time:.4f}s")
+        x = [np.transpose(x[i],(0,3,1,2)) for i in range(4)]    #to nchw
+        for i in range(len(x)):
+            reg_pred = x[i][:, :self.reg_max1 * 4,:,:]
+            cls_pred = x[i][:, self.reg_max1 * 4:,:,:]
+            out = np.concatenate([cls_pred, reg_pred], axis=1)
+            outputs.append(out.reshape(out.shape[0], out.shape[1], -1))
+        preds = np.concatenate(outputs, axis=2).transpose(0, 2, 1)
+
+        cls_scores = preds[:, :, :self.nc]
+        bbox_preds = preds[:, :, self.nc:]
+        pred = self.get_bboxes(cls_scores, bbox_preds)[0]
+        res = self.post_process(pred, org_data, im0, save_path, img_name)
+        result_img, bbox_res = overlay_bbox_cv(im0, res, names, score_thresh=0.35)
+        return bbox_res, result_img
+    def post_process(self, result, im, im0, save_path, img_name):
+        det_result = {}
+        det_bboxes, det_labels = result
+        det_bboxes[:, :4] = scale_boxes_no_letter(im.shape[:2], det_bboxes[:, :4], im0.shape).round()
+        classes = det_labels
+        for i in range(self.nc):
+            inds = classes == i
+            det_result[i] = np.concatenate(
+                [
+                    det_bboxes[inds, :4].astype(np.float32),
+                    det_bboxes[inds, 4:5].astype(np.float32),
+                ],
+                axis=1,
+            ).tolist()
+
+        return det_result
+    
+class QRCodeDecoder:
+    def crop_qr_regions(self, image, regions):
+        """
+        根据检测到的边界框裁剪二维码区域
+        """
+        cropped_images = []
+        for idx, region in enumerate(regions):
+            label, x1, y1, x2, y2, score = region
+            # 外扩15个像素缓解因检测截断造成无法识别的情况，视检测情况而定
+            x1-=15
+            y1-=15
+            x2+=15
+            y2+=15
+            # 裁剪图像
+            cropped = image[y1:y2, x1:x2]
+            if cropped.size > 0:
+                cropped_images.append({
+                    'image': cropped,
+                    'bbox': region,
+                })
+        return cropped_images
+
+    def decode_qrcode_pyzbar(self, cropped_image):
+        """
+        使用pyzbar解码二维码
+        """
+        try:
+            # 转换为灰度图像
+            if len(cropped_image.shape) == 3:
+                gray = cv2.cvtColor(cropped_image, cv2.COLOR_BGR2GRAY)
+            else:
+                gray = cropped_image
+            # 使用pyzbar解码
+            decoded_objects = pyzbar.decode(gray)
+            results = []
+            for obj in decoded_objects:
+                try:
+                    data = obj.data.decode('utf-8')
+                    results.append({
+                        'data': data,
+                        'type': obj.type,
+                        'points': obj.polygon
+                    })
+                except:
+                    continue
+            
+            return results
+        except Exception as e:
+            print(f"decode error: {e}")
+            return []
+
+if __name__ == '__main__':
+    import time
+
+    detector = NanoDetONNXInfer(model_path='./nanodet-plus-m_416_QR.onnx',imgsz=[416,416])
+    decoder = QRCodeDecoder()
+    img_path = './qrcode_test'
+    det_path='./det_res'
+    crop_path='./crop_res'
+    os.makedirs(det_path, exist_ok=True)
+    os.makedirs(crop_path, exist_ok=True)
+    imgs = glob.glob(f"{img_path}/*.jpg")
+    totoal = len(imgs)
+    success = 0
+    fail = 0
+    start_time = time.time()
+    for idx,img in enumerate(imgs):
+        pic_name=os.path.basename(img).split('.')[0]
+        loop_start_time = time.time()
+        det_result, res_img = detector.detect_objects(img,det_path)
+        # cv2.imwrite(os.path.join(det_path, pic_name+'.jpg'), res_img)
+        # print('det_result:',det_result)
+        # Crop deteted QRCode & decode QRCode by pyzbar
+        cropped_images = decoder.crop_qr_regions(res_img, det_result)
+        # for i,cropped in enumerate(cropped_images):
+        #     cv2.imwrite(os.path.join(crop_path, f'{pic_name}_crop_{i}.jpg'), cropped['image'])
+
+        all_decoded_results = []
+        for i, cropped_data in enumerate(cropped_images):
+            decoded_results = decoder.decode_qrcode_pyzbar(cropped_data['image'])
+            all_decoded_results.extend(decoded_results)
+            
+            # for result in decoded_results:
+            #     print(f"decode result: {result['data']} (type: {result['type']})")
+        if all_decoded_results:
+            success += 1
+            print(f"{pic_name} 识别成功！")
+        else:
+            fail += 1
+            print(f"{pic_name} 识别失败！")
+        loop_end_time = time.time()
+        print(f"图片 {img} 处理耗时: {loop_end_time - loop_start_time:.4f} 秒")
+
+    end_time = time.time()  # 记录总结束时间
+    total_time = end_time - start_time  # 记录总耗时
+
+    print(f"总共测试图片数量: {totoal}")
+    print(f"识别成功数量: {success}")
+    print(f"识别失败数量: {fail}")
+    print(f"识别成功率: {success/totoal*100:.2f}%")
+    print(f"整体处理耗时: {total_time:.4f} 秒")
+    print(f"平均每张图片处理耗时: {total_time/totoal:.4f} 秒")