app.py

from ultralytics import YOLO
import gradio as gr
import cv2
import numpy as np
from collections import defaultdict
import os
import torch

# 初始化模型
model = None

def load_model():
    global model
    if model is None:
        model = YOLO('./yolo11x-seg.pt')
    return model

def segment_image(image, conf_threshold, iou_threshold, mask_threshold, line_thickness, use_retina_masks):
    # 加载模型
    model = load_model()
    
    # 确保图像是BGR格式
    if len(image.shape) == 2:
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
    elif image.shape[2] == 4:
        image = cv2.cvtColor(image, cv2.COLOR_RGBA2BGR)
    
    # 使用模型进行预测
    results = model(
        image,
        conf=conf_threshold,
        iou=iou_threshold,
        device='cpu',
        retina_masks=use_retina_masks
    )
    result = results[0]
    
    # 按类别分组存储结果
    class_images = defaultdict(lambda: image.copy())
    detected_classes = set()
    
    if result.masks is not None:
        names = model.names
        # 处理每个检测结果
        for seg, box, cls in zip(result.masks, result.boxes, result.boxes.cls):
            class_id = int(cls)
            class_name = names[class_id]
            detected_classes.add(class_name)
            
            output_image = class_images[class_name]
            
            # 处理分割掩码
            segment = seg.data[0].cpu().numpy()
            segment = cv2.resize(segment, (output_image.shape[1], output_image.shape[0]))
            
            # 生成颜色
            color_mask = np.array([hash(class_name) % 256, 
                                 hash(class_name * 2) % 256,
                                 hash(class_name * 3) % 256], dtype=np.uint8)
            
            # 应用掩码
            mask_area = segment > mask_threshold
            overlay = output_image.copy()
            overlay[mask_area] = color_mask
            cv2.addWeighted(overlay, 0.4, output_image, 0.6, 0, output_image)
            
            # 添加边界框和标签
            conf = float(box.conf)
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            cv2.rectangle(output_image, (x1, y1), (x2, y2), 
                         color_mask.tolist(), line_thickness)
            
            # 添加标签
            label = f"{class_name} {conf:.2f}"
            font_scale = 0.6 * line_thickness / 2
            (label_width, label_height), _ = cv2.getTextSize(
                label, cv2.FONT_HERSHEY_SIMPLEX, font_scale, line_thickness)
            
            cv2.rectangle(output_image, 
                         (x1, y1 - label_height - 10), 
                         (x1 + label_width, y1), 
                         color_mask.tolist(), -1)
            
            cv2.putText(output_image, label, (x1, y1 - 5),
                       cv2.FONT_HERSHEY_SIMPLEX, font_scale, 
                       (255, 255, 255), line_thickness, cv2.LINE_AA)
            
            class_images[class_name] = output_image
    
    # 准备Gallery输出
    gallery_output = []
    
    # 添加完整结果
    if detected_classes:
        full_result = image.copy()
        for class_name in detected_classes:
            cv2.addWeighted(class_images[class_name], 0.5, full_result, 0.5, 0, full_result)
        gallery_output.append((full_result, "完整结果"))
    
        # 添加各个类别的结果
        for class_name in detected_classes:
            gallery_output.append((class_images[class_name], class_name))
    
    return gallery_output if gallery_output else None

def create_demo():
    with gr.Blocks() as demo:
        gr.Markdown("# YOLO 图像分割")
        gr.Markdown("上传一张图片，模型将对图片进行实例分割。每个检测到的类别将单独显示。")
        
        with gr.Row():
            with gr.Column(scale=1):
                input_image = gr.Image()
                with gr.Row():
                    conf_threshold = gr.Slider(
                        minimum=0.1, maximum=1.0, value=0.25, step=0.05,
                        label="置信度阈值", info="检测置信度的最小值"
                    )
                    iou_threshold = gr.Slider(
                        minimum=0.1, maximum=1.0, value=0.7, step=0.05,
                        label="IOU阈值", info="非极大值抑制的IOU阈值"
                    )
                with gr.Row():
                    mask_threshold = gr.Slider(
                        minimum=0.1, maximum=1.0, value=0.5, step=0.05,
                        label="掩码阈值", info="分割掩码的阈值"
                    )
                    line_thickness = gr.Slider(
                        minimum=1, maximum=5, value=2, step=1,
                        label="线条粗细", info="边界框和文本的粗细"
                    )
                with gr.Row():
                    retina_masks = gr.Checkbox(
                        label="高分辨率掩码",
                        value=True,
                        info="启用高分辨率分割掩码（可能会降低速度）"
                    )
            
            with gr.Column(scale=1):
                output_gallery = gr.Gallery(
                    label="分割结果",
                    show_label=True,
                    columns=2,
                    rows=2,
                    height=600,
                    object_fit="contain"
                )
        
        submit_btn = gr.Button("开始分割")
        
        submit_btn.click(
            fn=segment_image,
            inputs=[
                input_image,
                conf_threshold,
                iou_threshold,
                mask_threshold,
                line_thickness,
                retina_masks
            ],
            outputs=output_gallery
        )
    
    return demo

if __name__ == "__main__":
    demo = create_demo()
    demo.launch(share=True)