app.py

import gradio as gr
import torch
from qai_hub_models.models.detr_resnet50 import Model
from PIL import Image, ImageDraw
import numpy as np

# 注册 AVIF 支持（根据所安装的插件选择一种）
try:
    from pillow_avif import register_avif_opener
    register_avif_opener()
except ImportError:
    try:
        import pillow_heif
        pillow_heif.register_heif_opener()
    except ImportError:
        print("AVIF support not available. Please install 'pillow-avif-plugin' or 'pillow-heif'.")

# 加载模型
torch_model = Model.from_pretrained()

def detect_objects(image):
    if image is None:
        raise ValueError("No image uploaded!")  # 检查图像是否为 None

    # 将图像转换为 RGB 格式并调整大小
    image = Image.fromarray(image).convert("RGB")
    original_image = image.copy()  # 保存原始图像以便绘制边界框
    image = image.resize((800, 800))  # 根据模型要求调整图像大小

    # 转换为张量并进行标准化
    image_tensor = torch.tensor(np.array(image)).permute(2, 0, 1)  # 转换为 (C, H, W) 格式
    image_tensor = image_tensor.float() / 255.0  # 将像素值归一化到 [0, 1]
    
    # 使用模型进行推理
    with torch.no_grad():  # 禁用梯度计算以提高性能
        outputs = torch_model(image_tensor.unsqueeze(0))  # 添加批次维度

    # 获取预测结果（根据具体的输出格式进行调整）
    predictions = outputs['logits'] if 'logits' in outputs else outputs[0]  # 确保获取正确的输出

    # 格式化输出结果
    detections = []
    confidence_threshold = 0.8  # 设置置信度阈值为 0.8
    
    for i in range(predictions.shape[1]):  # 遍历每个预测
        score = predictions[0, i, -1].item()  # 假设最后一个维度是分数
        if score > confidence_threshold:  # 使用新的阈值过滤低置信度的预测
            
            box = predictions[0, i, :-1].tolist()  # 获取边界框坐标（假设在前面）
            box[0] *= original_image.width / 800  # 将坐标缩放回原始图像尺寸
            box[1] *= original_image.height / 800
            box[2] *= original_image.width / 800
            box[3] *= original_image.height / 800
            
            detections.append({
                "label": f"Object {i}",  
                "confidence": round(score, 3),
                "box": box,
            })
            
            # 绘制边界框和标签到原始图像上
            draw = ImageDraw.Draw(original_image)
            if box[1] < box[3]:  # 确保 y_min < y_max
                draw.rectangle(box[:4], outline="red", width=3)  
                draw.text((box[0], box[1]), f"{detections[-1]['label']} ({detections[-1]['confidence']})", fill="red")  

    return original_image, detections

# 创建 Gradio 接口，自动处理 AVIF 图像并转换为 PNG 格式以供显示和处理。
with gr.Blocks() as iface:
    gr.Markdown("# Object Detection with DETR-ResNet50")
    
    with gr.Row():
        with gr.Column(scale=1):
            image_input = gr.Image(type="numpy", label="Upload Image (supports PNG, JPEG, AVIF...)")  
            submit_button = gr.Button("Submit")
            clear_button = gr.Button("Clear")
        with gr.Column(scale=1):
            output_image = gr.Image(label="Detected Image")
            output_json = gr.JSON(label="Detections")

    def on_submit(image):
        try:
            detected_image, detections = detect_objects(image)
            return detected_image, detections
        except Exception as e:
            return None, {"error": str(e)}

    def on_clear():
        return None, None, None  # 清空输入和输出

    submit_button.click(on_submit, inputs=image_input, outputs=[output_image, output_json])
    clear_button.click(on_clear, inputs=None, outputs=[image_input, output_image, output_json])  # 修复清除功能

# 启动应用程序
if __name__ == "__main__":
    iface.launch()