import io import gradio as gr import matplotlib.pyplot as plt import requests, validators import torch import pathlib from PIL import Image from transformers import AutoFeatureExtractor, YolosForObjectDetection, DetrForObjectDetection import os os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" # colors for visualization COLORS = [ [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] ] import numpy as np import tensorflow as tf # Load EV plate classifier ev_model = tf.keras.models.load_model("plate_color_model.h5") def is_green_plate(plate_img): plate_img = plate_img.resize((128,128)) plate_img = np.array(plate_img)/255.0 plate_img = np.expand_dims(plate_img, axis=0) pred = ev_model.predict(plate_img)[0][0] return pred > 0.5 def make_prediction(img, feature_extractor, model): inputs = feature_extractor(img, return_tensors="pt") outputs = model(**inputs) img_size = torch.tensor([tuple(reversed(img.size))]) processed_outputs = feature_extractor.post_process(outputs, img_size) return processed_outputs[0] def fig2img(fig): buf = io.BytesIO() fig.savefig(buf) buf.seek(0) pil_img = Image.open(buf) basewidth = 750 wpercent = (basewidth/float(pil_img.size[0])) hsize = int((float(pil_img.size[1])*float(wpercent))) img = pil_img.resize((basewidth,hsize), Image.Resampling.LANCZOS) return img def visualize_prediction(img, output_dict, threshold=0.5, id2label=None): boxes = output_dict["boxes"].tolist() scores = output_dict["scores"].tolist() labels = output_dict["labels"].tolist() if id2label is not None: labels = [id2label[x] for x in labels] plt.figure(figsize=(20, 20)) plt.imshow(img) ax = plt.gca() colors = COLORS * 100 for score, (xmin, ymin, xmax, ymax), label, color in zip(scores, boxes, labels, colors): if score < threshold: continue if label in ["license-plates", "Rego Plates"]: plate_crop = img.crop((xmin, ymin, xmax, ymax)) ev = is_green_plate(plate_crop) if ev: plate_type = "EV (Green Plate)" box_color = "green" else: plate_type = "Non-EV Plate" box_color = "red" else: plate_type = label box_color = "blue" ax.add_patch( plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, fill=False, color=box_color, linewidth=3) ) ax.text( xmin, ymin - 10, f"{plate_type} | {score:.2f}", fontsize=14, bbox=dict(facecolor=box_color, alpha=0.7), color="white" ) plt.axis("off") return fig2img(plt.gcf()) def get_original_image(url_input): if validators.url(url_input): image = Image.open(requests.get(url_input, stream=True).raw) return image def detect_objects(model_name,url_input,image_input,webcam_input,threshold): #Extract model and feature extractor feature_extractor = AutoFeatureExtractor.from_pretrained(model_name) if "yolos" in model_name: model = YolosForObjectDetection.from_pretrained(model_name) elif "detr" in model_name: model = DetrForObjectDetection.from_pretrained(model_name) if validators.url(url_input): image = get_original_image(url_input) elif image_input: image = image_input elif webcam_input: image = webcam_input #Make prediction processed_outputs = make_prediction(image.convert("RGB"), feature_extractor, model) #Visualize prediction viz_img = visualize_prediction(image, processed_outputs, threshold, model.config.id2label) return viz_img def set_example_image(example: list) -> dict: return gr.Image.update(value=example[0]) def set_example_url(example: list) -> dict: return gr.Textbox.update(value=example[0]), gr.Image.update(value=get_original_image(example[0])) title = """

License Plate Detection with YOLOS

""" description = """ YOLOS is a Vision Transformer (ViT) trained using the DETR loss. Despite its simplicity, a base-sized YOLOS model is able to achieve 42 AP on COCO validation 2017 (similar to DETR and more complex frameworks such as Faster R-CNN). The YOLOS model was fine-tuned on COCO 2017 object detection (118k annotated images). It was introduced in the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Fang et al. and first released in [this repository](https://github.com/hustvl/YOLOS). This model was further fine-tuned on the [Car license plate dataset]("https://www.kaggle.com/datasets/andrewmvd/car-plate-detection") from Kaggle. The dataset consists of 443 images of vehicle with annotations categorised as "Vehicle" and "Rego Plates". The model was trained for 200 epochs on a single GPU. Links to HuggingFace Models: - [nickmuchi/yolos-small-rego-plates-detection](https://huggingface.co/nickmuchi/yolos-small-rego-plates-detection) - [hustlv/yolos-small](https://huggingface.co/hustlv/yolos-small) """ models = ["nickmuchi/yolos-small-finetuned-license-plate-detection","nickmuchi/detr-resnet50-license-plate-detection"] urls = ["https://drive.google.com/uc?id=1j9VZQ4NDS4gsubFf3m2qQoTMWLk552bQ","https://drive.google.com/uc?id=1p9wJIqRz3W50e2f_A0D8ftla8hoXz4T5"] images = [[path.as_posix()] for path in sorted(pathlib.Path('images').rglob('*.j*g'))] twitter_link = """ [![](https://img.shields.io/twitter/follow/nickmuchi?label=@nickmuchi&style=social)](https://twitter.com/nickmuchi) """ css = ''' h1#title { text-align: center; } ''' demo = gr.Blocks() with demo: gr.Markdown(title) gr.Markdown(description) gr.Markdown(twitter_link) options = gr.Dropdown(choices=models,label='Object Detection Model',value=models[0],show_label=True) slider_input = gr.Slider(minimum=0.2,maximum=1,value=0.5,step=0.1,label='Prediction Threshold') with gr.Tabs(): with gr.TabItem('Image URL'): with gr.Row(): with gr.Column(): url_input = gr.Textbox(lines=2,label='Enter valid image URL here..') original_image = gr.Image(height=750, width=750) url_input.change(get_original_image, url_input, original_image) with gr.Column(): img_output_from_url = gr.Image(height=750, width=750) with gr.Row(): example_url = gr.Examples(examples=urls,inputs=[url_input]) url_but = gr.Button('Detect') with gr.TabItem('Image Upload'): with gr.Row(): img_input = gr.Image(type='pil',height=750, width=750) img_output_from_upload= gr.Image(height=750, width=750) with gr.Row(): example_images = gr.Examples(examples=images,inputs=[img_input]) img_but = gr.Button('Detect') with gr.TabItem('WebCam'): with gr.Row(): web_input = gr.Image(type="pil", height=750, width=750, sources=["webcam"]) img_output_from_webcam= gr.Image(height=750, width=750) cam_but = gr.Button('Detect') url_but.click(detect_objects,inputs=[options,url_input,img_input,web_input,slider_input],outputs=[img_output_from_url],queue=True) img_but.click(detect_objects,inputs=[options,url_input,img_input,web_input,slider_input],outputs=[img_output_from_upload],queue=True) cam_but.click(detect_objects,inputs=[options,url_input,img_input,web_input,slider_input],outputs=[img_output_from_webcam],queue=True) gr.Markdown("![visitor badge](https://visitor-badge.glitch.me/badge?page_id=nickmuchi-license-plate-detection-with-yolos)") demo.launch(debug=True, css=css)