File size: 7,293 Bytes
0961b15 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 | #!/usr/bin/env python
"""
ONNX 单图推理脚本 (不依赖 PyTorch)。
功能：读取单张图片 → 预处理到模型尺寸 → 推理 → 还原到原图尺寸 → 保存「原图 + 去噪结果」横向拼接图。
- 原图 ≤ 模型尺寸：reflect pad 到模型尺寸，推理后裁回原图
- 原图 > 模型尺寸：resize 到模型尺寸，推理后 resize 回原图
所有参数均以固定默认值写在脚本顶部，直接运行即可。
"""
import os
import sys
import cv2
import numpy as np
# ═══════════════════════════════════════════════════════════════
# 固定默认参数（按需修改）
# ═══════════════════════════════════════════════════════════════
DEFAULT_ONNX_MODEL = os.path.join("onnx_models", "real_denoising_224x224_sim.onnx")
DEFAULT_INPUT_IMAGE = os.path.join("demo", "mixed/noisy.png")
DEFAULT_OUTPUT_IMAGE = "onnx_result_compare.png"
# ONNX 模型导出时的固定输入尺寸 (H, W)，必须能被 8 整除
MODEL_HEIGHT = 224
MODEL_WIDTH = 224
# 推理后端：cuda / cpu
PROVIDER = "cuda" # 可选: "cuda", "cpu"
# 是否保存纯去噪结果（不拼接）
SAVE_RESTORED_ONLY = False
RESTORED_ONLY_PATH = "onnx_result_restored.png"
# ═══════════════════════════════════════════════════════════════
def _load_image(path):
"""用 OpenCV 读取图片，返回 RGB uint8 numpy (H,W,3) 或 (H,W,1)。"""
if not os.path.isfile(path):
raise FileNotFoundError(f"输入图片不存在: {path}")
img = cv2.imread(path, cv2.IMREAD_COLOR)
if img is None:
img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
if img is None:
raise RuntimeError(f"无法读取图片: {path}")
return img[:, :, None] # (H,W,1)
return cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
def _save_image(path, img):
"""保存 uint8 numpy 图片，自动处理灰度/彩色。"""
os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
if img.ndim == 2:
cv2.imwrite(path, img)
elif img.shape[2] == 1:
cv2.imwrite(path, img[:, :, 0])
else:
cv2.imwrite(path, cv2.cvtColor(img, cv2.COLOR_RGB2BGR))
def _image_to_nchw(img):
"""uint8 (H,W,C) → float32 NCHW [0,1]。"""
return np.transpose(img.astype(np.float32) / 255.0, (2, 0, 1))[np.newaxis, ...]
def _nchw_to_image(arr):
"""float32 NCHW [0,1] → uint8 (H,W,C)。"""
arr = arr.squeeze(0)
arr = np.transpose(arr, (1, 2, 0))
arr = np.clip(arr, 0.0, 1.0)
return (arr * 255.0).round().clip(0, 255).astype(np.uint8)
def _prepare_for_model(tensor, model_h, model_w):
"""
将 NCHW tensor 处理到模型固定尺寸：
- 原图 ≤ 模型尺寸：reflect pad → 记录原始区域 → 返回 (padded, (crop_h, crop_w), mode="pad")
- 原图 > 模型尺寸：resize → 返回 (resized, (orig_h, orig_w), mode="resize")
"""
_, c, orig_h, orig_w = tensor.shape
if orig_h <= model_h and orig_w <= model_w:
# 小图/等尺寸：reflect pad 到模型尺寸
pad_h = model_h - orig_h
pad_w = model_w - orig_w
if pad_h or pad_w:
tensor = np.pad(tensor, ((0, 0), (0, 0), (0, pad_h), (0, pad_w)), mode="reflect")
return tensor.astype(np.float32), (orig_h, orig_w), "pad"
else:
# 大图：resize 到模型尺寸
resized = np.zeros((1, c, model_h, model_w), dtype=np.float32)
for b in range(tensor.shape[0]):
for ch in range(c):
resized[b, ch] = cv2.resize(tensor[b, ch], (model_w, model_h), interpolation=cv2.INTER_LINEAR)
return resized.astype(np.float32), (orig_h, orig_w), "resize"
def _restore_from_model(output_arr, orig_h, orig_w, mode):
"""
将模型输出还原到原图尺寸：
- mode="pad"：裁掉 padding 区域
- mode="resize"：resize 回原图尺寸
"""
if mode == "pad":
return output_arr[:, :, :orig_h, :orig_w]
else:
_, c, _, _ = output_arr.shape
restored = np.zeros((1, c, orig_h, orig_w), dtype=np.float32)
for b in range(output_arr.shape[0]):
for ch in range(c):
restored[b, ch] = cv2.resize(output_arr[b, ch], (orig_w, orig_h), interpolation=cv2.INTER_LINEAR)
return restored
def main():
# ----- 命令行覆盖参数（可选） -----
onnx_path = sys.argv[1] if len(sys.argv) > 1 else DEFAULT_ONNX_MODEL
input_path = sys.argv[2] if len(sys.argv) > 2 else DEFAULT_INPUT_IMAGE
output_path = sys.argv[3] if len(sys.argv) > 3 else DEFAULT_OUTPUT_IMAGE
if MODEL_HEIGHT % 8 != 0 or MODEL_WIDTH % 8 != 0:
raise ValueError(f"MODEL_HEIGHT({MODEL_HEIGHT}) 和 MODEL_WIDTH({MODEL_WIDTH}) 必须能被 8 整除")
# ----- 加载 ONNX Runtime -----
try:
import onnxruntime as ort
except ImportError:
raise ImportError("请先安装 onnxruntime: pip install onnxruntime-gpu 或 pip install onnxruntime")
if PROVIDER == "cuda":
providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
else:
providers = ["CPUExecutionProvider"]
session = ort.InferenceSession(onnx_path, providers=providers)
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name
# ----- 读取原图 -----
original_img = _load_image(input_path) # uint8 HWC RGB
orig_h, orig_w = original_img.shape[:2]
is_gray = original_img.ndim == 2 or original_img.shape[2] == 1
# ----- 预处理到模型尺寸 -----
tensor = _image_to_nchw(original_img)
tensor, (crop_h, crop_w), prep_mode = _prepare_for_model(tensor, MODEL_HEIGHT, MODEL_WIDTH)
print(f"[ONNX 推理] 模型: {onnx_path}")
print(f"[ONNX 推理] 输入图片: {input_path} (原图尺寸: {orig_h}x{orig_w}, 模型输入尺寸: {MODEL_HEIGHT}x{MODEL_WIDTH}, 预处理模式: {prep_mode})")
# ----- ONNX 推理 -----
ort_output = session.run([output_name], {input_name: tensor})[0]
# ----- 后处理：还原到原图尺寸 -----
restored_tensor = _restore_from_model(ort_output, crop_h, crop_w, prep_mode)
restored_img = _nchw_to_image(restored_tensor) # uint8 HWC RGB, 尺寸 = (crop_h, crop_w)
# ----- 拼接：原图 || 去噪结果 -----
if is_gray:
orig_display = original_img[:, :, 0] if original_img.ndim == 3 else original_img
rest_display = restored_img[:, :, 0] if restored_img.ndim == 3 else restored_img
comparison = np.concatenate([orig_display, rest_display], axis=1)
else:
comparison = np.concatenate([original_img, restored_img], axis=1)
# ----- 保存结果 -----
_save_image(output_path, comparison)
print(f"[ONNX 推理] 拼接结果已保存到: {output_path} (左=原图, 右=去噪)")
if SAVE_RESTORED_ONLY:
_save_image(RESTORED_ONLY_PATH, restored_img)
print(f"[ONNX 推理] 纯去噪结果已保存到: {RESTORED_ONLY_PATH}")
if __name__ == "__main__":
main()
|