自动通过腾讯点选验证码。golang + gocv + yolov5

go-yolov5-onnx.go

package main

import (
	"fmt"
	"image"
	"image/color"

	"gocv.io/x/gocv"
)

// 定义一些常量
const (
	confidenceThreshold float32 = 0.5 // 置信度阈值
	nmsThreshold        float32 = 0.3 // NMS 阈值

	// 模型训练的类别数
	numClasses = 1
)

func main() {
	// 1. 加载 ONNX 模型
	onnxModelPath := "./best.onnx"
	net := gocv.ReadNetFromONNX(onnxModelPath)
	if net.Empty() {
		fmt.Printf("Error reading network model from: %v\n", onnxModelPath)
		return
	}
	defer net.Close()

	// 2. 读取并预处理输入图片
	imagePath := "./target_bg_image.png"
	img := gocv.IMRead(imagePath, gocv.IMReadColor)
	if img.Empty() {
		fmt.Printf("Error reading image from: %s\n", imagePath)
		return
	}
	defer img.Close()

	blob := gocv.BlobFromImage(
		img,
		1.0/255.0,
		image.Pt(672, 480),
		gocv.NewScalar(0, 0, 0, 0),
		true, // 交换RB通道
		false,
	)
	defer blob.Close()

	// 3. 将处理好的图片输入网络并执行推理
	net.SetInput(blob, "")

	layerNames := net.GetLayerNames()
	outLayerIndices := net.GetUnconnectedOutLayers()
	outputLayerName := layerNames[outLayerIndices[0]-1]
	outputs := net.Forward(outputLayerName)

	if outputs.Empty() {
		fmt.Println("Error: Forward pass returned no outputs.")
		return
	}
	defer outputs.Close()

	// 4. 解析和后处理输出结果
	fmt.Printf("Raw output matrix info: Rows=%d, Cols=%d, Total=%d\n", outputs.Rows(), outputs.Cols(), outputs.Total())

	// ======================== 关键修正点 ========================
	// 手动重塑矩阵，因为 gocv.Forward() 未能正确解析输出维度
	numCols := 5 + numClasses
	numRows := outputs.Total() / numCols

	// 将一维的Mat重塑为 [numRows, numCols] 的二维Mat
	reshaped := outputs.Reshape(1, numRows)
	defer reshaped.Close() // reshaped是一个新的Mat头，需要独立关闭

	fmt.Printf("Reshaped matrix: Rows=%d, Cols=%d\n", reshaped.Rows(), reshaped.Cols())
	// ==========================================================

	// 使用重塑后的矩阵进行后处理
	boxes, confidences, classIDs := postProcess(reshaped, img.Cols(), img.Rows())

	if len(boxes) == 0 {
		fmt.Println("No objects detected after post-processing.")
		return
	}

	indices := gocv.NMSBoxes(boxes, confidences, confidenceThreshold, nmsThreshold)
	if len(indices) == 0 {
		fmt.Println("No objects detected after NMS.")
		return
	}

	fmt.Printf("Detected %d objects:\n", len(indices))
	for _, i := range indices {
		box := boxes[i]
		classID := classIDs[i]
		confidence := confidences[i]
		fmt.Printf("  - Class ID: %d, Confidence: %.2f, BBox: %v\n", classID, confidence, box)

		green := color.RGBA{0, 255, 0, 255}
		gocv.Rectangle(&img, box, green, 2)
		gocv.PutText(&img, fmt.Sprintf("Class %d", classID), image.Pt(box.Min.X, box.Min.Y-5), gocv.FontHersheySimplex, 0.5, green, 2)
	}

	gocv.IMWrite("result.jpg", img)
	fmt.Println("Result image saved to result.jpg")
}

// postProcess 函数现在接收重塑后的矩阵，其内部逻辑无需修改
func postProcess(output gocv.Mat, originalImgWidth, originalImgHeight int) ([]image.Rectangle, []float32, []int) {
	var boxes []image.Rectangle
	var confidences []float32
	var classIDs []int

	rows := output.Rows()
	cols := output.Cols()

	xScale := float32(originalImgWidth) / 672.0
	yScale := float32(originalImgHeight) / 480.0

	for r := 0; r < rows; r++ {
		confidence := output.GetFloatAt(r, 4)
		if confidence < confidenceThreshold {
			continue
		}

		var maxScore float32 = 0.0
		var classID int = 0
		// 从第5列开始是类别得分
		for c := 5; c < cols; c++ {
			score := output.GetFloatAt(r, c)
			if score > maxScore {
				maxScore = score
				classID = c - 5
			}
		}

		if maxScore*confidence > confidenceThreshold {
			confidences = append(confidences, confidence)
			classIDs = append(classIDs, classID)

			cx := output.GetFloatAt(r, 0) * xScale
			cy := output.GetFloatAt(r, 1) * yScale
			w := output.GetFloatAt(r, 2) * xScale
			h := output.GetFloatAt(r, 3) * yScale

			x1 := int(cx - w/2)
			y1 := int(cy - h/2)
			x2 := int(cx + w/2)
			y2 := int(cy + h/2)

			boxes = append(boxes, image.Rect(x1, y1, x2, y2))
		}
	}

	return boxes, confidences, classIDs
}

windows 环境中，配置 gocv 环境是个天坑

安装 scoop 后跟随教程 ：
https://www.cnblogs.com/beidaxmf/p/18258953

• 此时你应当已经解决 mingw、cmake 环境问题。
• 接下来解决 gocv 环境问题

若 golang 版本高于 1.13 大概率是无法运行 win_build_opencv.cmd，参考解决办法：
https://www.cnblogs.com/liulog/p/18822802

若是 cmd 运行遇到问题

.... Compatibility with CMake < 3.5 has been removed from CMake. Update the VERSION argument value. ....

可看：

https://www.cnblogs.com/liulog/p/18822802

https://github.com/hybridgroup/gocv/issues/1320#issuecomment-2993421780

项目备份

源码留存备份：

• golang源码下载

训练数据：

• 数据标注平台
• 已标注数据下载

路线图：

• 继续跟随 https://blog.2zxz.com/archives/icondetection 完成 Simaese孪生神经网络 部分。

运行

yolov5 已经被新的 yolo11 取代。

yolotool 已开始删除对 yolov5 的支持

在 yolov5 中常用的命令：

激活 venv

.venv/scripts/activate

开始训练

python train.py --workers 0

将权重 *.pt 文件转为通用 onnx 模型

python export.py --weights runs/train/exp/weights/best.pt