X-AnyLabeling项目中使用自定义模型进行半自动标注的技术指南

2025-06-07 11:56:46作者：冯爽妲Honey

前言

在计算机视觉领域，标注数据是模型训练的重要基础工作。X-AnyLabeling作为一款优秀的标注工具，不仅提供了丰富的预训练模型支持，还允许用户导入自定义模型进行半自动标注，这大大提高了标注效率。本文将详细介绍如何在X-AnyLabeling中使用自定义模型进行半自动标注。

自定义模型的基本流程

使用自定义模型进行半自动标注主要包含以下几个步骤：

模型训练：在自己的数据集上训练模型
模型转换：将训练好的模型转换为ONNX格式
模型集成：将转换后的模型集成到X-AnyLabeling中
模型测试：验证模型在标注工具中的效果

模型转换详解

X-AnyLabeling要求模型必须为ONNX格式。对于PyTorch训练的用户，可以使用以下代码示例进行转换：

import torch
import segmentation_models_pytorch as smp

def pth_to_onnx():
    # 加载训练好的模型
    torch_model = torch.load('model.pth')
    model = smp.UnetPlusPlus(
        encoder_name="timm-regnety_320",
        encoder_weights="imagenet",
        in_channels=3,
        classes=2,
    )
    model.load_state_dict(torch_model)
    
    # 设置模型为评估模式
    model.eval()
    
    # 定义输入张量
    batch_size = 32
    input_shape = (3, 256, 256)
    x = torch.randn(batch_size, *input_shape)
    
    # 导出ONNX模型
    torch.onnx.export(
        model,
        x,
        "model.onnx",
        opset_version=11,
        do_constant_folding=True,
        input_names=["input"],
        output_names=["output"],
        dynamic_axes={
            "input": {0: "batch_size"},
            "output": {0: "batch_size"}
        }
    )

转换时需要注意以下几点：

确保opset_version与X-AnyLabeling兼容
检查输入输出名称是否正确
验证转换后的模型能否正常推理

自定义模型集成

在X-AnyLabeling中集成自定义模型需要创建一个继承自Model类的Python文件。以下是一个完整的示例：

import cv2
import numpy as np
from PyQt5.QtCore import QCoreApplication
from anylabeling.views.labeling.shape import Shape
from anylabeling.views.labeling.utils.opencv import qt_img_to_rgb_cv_img
from .model import Model
from .types import AutoLabelingResult
from .engines.build_onnx_engine import OnnxBaseModel

class CustomModel(Model):
    """自定义语义分割模型"""
    
    class Meta:
        required_config_names = [
            "type",
            "name",
            "display_name",
            "model_path",
            "classes",
        ]
        widgets = ["button_run"]
        output_modes = {
            "polygon": QCoreApplication.translate("Model", "多边形"),
        }
        default_output_mode = "polygon"

    def __init__(self, model_config, on_message):
        super().__init__(model_config, on_message)
        model_abs_path = self.get_model_abs_path(self.config, "model_path")
        self.net = OnnxBaseModel(model_abs_path, "cuda")
        self.classes = self.config["classes"]
        self.input_shape = self.net.get_input_shape()[-2:]

    def preprocess(self, input_image):
        """图像预处理"""
        input_h, input_w = self.input_shape
        image = cv2.resize(input_image, (input_w, input_h))
        image = np.transpose(image, (2, 0, 1))
        image = image.astype(np.float32) / 255.0
        image = (image - 0.5) / 0.5
        return np.expand_dims(image, axis=0)

    def postprocess(self, image, outputs):
        """结果后处理"""
        outputs = np.argmax(outputs, axis=1)
        results = []
        for i, class_name in enumerate(self.classes):
            if class_name == '_background_':
                continue
            mask = outputs[0] == i
            mask_resized = cv2.resize(
                mask.astype(np.uint8), 
                (image.shape[1], image.shape[0])
            contours, _ = cv2.findContours(
                mask_resized, 
                cv2.RETR_EXTERNAL, 
                cv2.CHAIN_APPROX_SIMPLE)
            results.append((
                class_name, 
                [np.squeeze(contour).tolist() for contour in contours]))
        return results

    def predict_shapes(self, image, image_path=None):
        """预测形状"""
        if image is None:
            return []
            
        try:
            image = qt_img_to_rgb_cv_img(image, image_path)
            blob = self.preprocess(image)
            outputs = self.net.get_ort_inference(blob)
            results = self.postprocess(image, outputs)
            
            shapes = []
            for label, contours in results:
                for points in contours:
                    points.append(points[0])  # 闭合多边形
                    shape = Shape(flags={})
                    for point in points:
                        shape.add_point(QtCore.QPointF(point[0], point[1]))
                    shape.shape_type = "polygon"
                    shape.closed = True
                    shape.label = label
                    shapes.append(shape)
                    
            return AutoLabelingResult(shapes, replace=True)
            
        except Exception as e:
            print(f"推理错误: {e}")
            return []