网上找了一个opencv_aardio库(https://github.com/xuncv/opencv_aardio)，折腾了半天，才配置好环境，为了更好的测试效果，对原库部分代码也进行了一些修改😏。
opencv确实强大，试了一下YOLO-tiny模型，虽然识别的准确度不高（可能因为版本问题，这个库目前不支持onnx），但还是挺惊艳的！~~
现在，完整cv2库文件也分享给大家，大家可以少走弯路(坑我已经踩过了😂)，直接import调用即可，我们一起探索opencv的世界吧：
【点击下载】：cv2.7z (访问密码: 1345)
（下载后，解压，放在arrdio文件夹的"\lib\"目录下就可以用了）
// opencv + YOLOv4-tiny模型 识别照片中的人物 @Mr_MAO
import win.ui;
import cv2;
import fsys.dlg;
/*DSG{{*/
var winform = win.form(text="YOLO 人体检测工具";right=899;bottom=599;bgcolor=0x2E1A1A)
winform.add(
btnDetect={cls="plus";text="检测人体";left=761;top=17;right=881;bottom=49;bgcolor=0x5EC522;border={radius=6};color=0xFFFFFF;disabled=1;dr=1;dt=1;font=LOGFONT(h=-14);notify=1;z=2};
btnOpen={cls="plus";text="选择图片";left=631;top=17;right=751;bottom=49;bgcolor=0xF76C4A;border={radius=6};color=0xFFFFFF;dr=1;dt=1;font=LOGFONT(h=-14);notify=1;z=1};
imageView={cls="plus";left=21;top=70;right=881;bottom=581;bgcolor=0x1A0F0F;border={color=0x48372D;radius=8;width=1};db=1;dl=1;dr=1;dt=1;repeat="scale";z=3};
infoOverlay={cls="bkplus";left=32;top=80;right=383;bottom=119;align="left";color=0x5EC522;dl=1;dt=1;font=LOGFONT(h=-24;weight=700);z=4};
statusText={cls="bkplus";text="YOLO 图像检测";left=25;top=20;right=599;bottom=48;color=0xC0AEA0;dl=1;dr=1;dt=1;font=LOGFONT(h=-14);z=5}
)
/*}}*/
// ====== 界面样式 ======
winform.btnOpen.skin({ background={default=0xFF4A6CF7;hover=0xFF5B7CF8;active=0xFF3A5CE7} })
winform.btnDetect.skin({
background={default=0xFF22C55E;hover=0xFF32D56E;active=0xFF18B54E;disabled=0xFF555555}
})
// ====== 全局配置与状态 ======
var modelDir = io.fullpath("~/lib/cv2/.res/model/");
io.createDir(modelDir);
// YOLOv4-tiny Darknet
var cfgPath = modelDir + "yolov4-tiny.cfg";
var weightsPath = modelDir + "yolov4-tiny.weights";
var currentMat = null; // 当前图片 Mat
var globalNet = null; // 缓存的 DNN 网络模型
// ====== 函数：显示图片 ======
var updateDisplay = function(mat){
if(!mat || mat.empty()) return;
var bgra = cv2.cvtColor(mat, cv2.COLOR_BGR2BGRA);
var bytes = bgra.toBytes();
var bmp = gdip.bitmap(bgra.cols, bgra.rows, 0x26200A, bytes, bgra.cols * 4);
winform.imageView.background = bmp;
bgra.release();
}
// ====== 函数：加载 YOLOv4-tiny Darknet 模型 ======
var loadNetwork = function(){
if(globalNet) return true;
if(!io.exist(cfgPath) || !io.exist(weightsPath)){
winform.statusText.text = "模型文件不存在，请等待模型准备完成";
winform.statusText.redraw();
return false;
}
var cfgSize = io.getSize(cfgPath);
var weightsSize = io.getSize(weightsPath);
if(cfgSize < 100 || weightsSize < 5000000){
winform.statusText.text = "模型文件不完整，正在重新准备...";
winform.statusText.redraw();
io.remove(cfgPath);
io.remove(weightsPath);
return false;
}
winform.statusText.text = "正在载入 AI 模型到内存...";
winform.statusText.redraw();
var net = null;
try {
net = cv2.readNetFromDarknet(cfgPath, weightsPath);
}
catch(e){
winform.statusText.text = "模型加载失败: " + (e || "未知错误");
winform.statusText.redraw();
return false;
}
if(!net || net.empty()){
winform.statusText.text = "模型文件损坏，正在重新准备...";
winform.statusText.redraw();
io.remove(cfgPath);
io.remove(weightsPath);
return false;
}
net.setPreferableBackend(cv2.DNN_BACKEND_OPENCV);
net.setPreferableTarget(cv2.DNN_TARGET_CPU);
globalNet = net;
winform.statusText.text = "模型加载成功，可以开始检测";
winform.statusText.redraw();
winform.btnDetect.disabled = false;
return true;
}
// ====== 函数：后台下载模型 ======
var downloadModel = function(){
thread.invoke(
function(winform, cfgPath, weightsPath, modelDir){
import web.rest.client;
var http = web.rest.client();
// YOLOv4-tiny cfg
var cfgUrl = "https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg";
// YOLOv4-tiny weights
var weightsUrl = "https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights";
winform.statusText.text = "正在下载 YOLOv4-tiny 模型 (约 23MB)...";
// 下载 cfg
var ok = http.receiveFile(cfgPath).get(cfgUrl);
if(!ok || io.getSize(cfgPath) < 100){
winform.statusText.text = "下载 cfg 失败，请检查网络连接";
return;
}
// 下载 weights
ok = http.receiveFile(weightsPath).get(weightsUrl);
if(!ok || io.getSize(weightsPath) < 5000000){
winform.statusText.text = "下载 weights 失败，请检查网络连接";
return;
}
winform.statusText.text = "模型下载完成，正在加载...";
}, winform, cfgPath, weightsPath, modelDir
);
};
// ====== 函数：模型准备入口 ======
var prepareModel = function(){
if(io.exist(cfgPath) && io.exist(weightsPath)
&& io.getSize(cfgPath) > 100 && io.getSize(weightsPath) > 5000000){
loadNetwork();
} else {
winform.statusText.text = "模型未就绪，开始在后台下载...";
winform.statusText.redraw();
downloadModel();
winform.setInterval(
function(){
if(io.exist(cfgPath) && io.exist(weightsPath)
&& io.getSize(cfgPath) > 100 && io.getSize(weightsPath) > 5000000){
if(loadNetwork()){
return false; //停止定时器
}
}
}, 2000
);
}
};
// ====== 函数：YOLOv4-tiny 检测 ======
var yoloDetect = function(net, mat, confThreshold=0.35, nmsThreshold=0.45){
var imgW = mat.cols;
var imgH = mat.rows;
// 创建 blob (YOLOv4-tiny 使用 416x416)
var blob = cv2.blobFromImage(mat, 1.0/255.0, [416, 416], null, true, false);
net.setInput(blob);
// 获取所有输出层名称并推理
var outNames = net.getUnconnectedOutLayersNames();
// 收集所有检测行
var allRows = [];
for(n, name in outNames){
var out = net.forward(name);
var bytes = out.toBytes();
var floatCount = #bytes / 4;
var cols = out.cols;
var rows = out.rows;
// 使用 raw.convert 批量将字节转为 float 数组
var st = raw.convert(bytes, {
float values[] = [length=floatCount];
});
// 按行组装
for(r=0; rows-1; 1){
var baseIdx = r * cols;
var row = [];
for(c=0; cols-1; 1){
table.push(row, st.values[baseIdx + c + 1]);
}
table.push(allRows, row);
}
out.release();
}
blob.release();
return cv2.dnnDecodeYolo(allRows, [imgW, imgH], confThreshold, nmsThreshold);
};
// ====== 函数：推理检测逻辑 ======
var doDetection = function(){
if(!currentMat || currentMat.empty()) return;
if(!globalNet){
if(!loadNetwork()){
winform.statusText.text = "模型未就绪，请等待模型下载完成";
winform.statusText.redraw();
return;
}
}
var startTick = time.tick();
winform.statusText.text = "AI 正在检测中...";
winform.statusText.redraw();
var detections = yoloDetect(globalNet, currentMat, 0.35, 0.45);
var renderMat = currentMat.clone();
var personCount = 0;
for(i, det in detections){
if(det.classId == 1){
personCount++;
var x = math.max(0, det.x);
var y = math.max(0, det.y);
cv2.rectangle(renderMat, [x, y], [x + det.width, y + det.height], [0, 255, 0], 2);
var labelY = math.max(20, y - 8);
cv2.putText(renderMat, "Person " + string.format("%.0f%%", det.score * 100),
[x, labelY], cv2.FONT_HERSHEY_SIMPLEX, 0.5, [0, 255, 0], 1);
}
}
updateDisplay(renderMat);
renderMat.release();
winform.infoOverlay.text = personCount > 0 ? "检测到 " + personCount + " 人" : "未检测到人";
winform.statusText.text = "检测完成，耗时: " + (time.tick() - startTick) + "ms";
winform.statusText.redraw();
winform.infoOverlay.redraw();
};
// ====== 交互事件 ======
winform.btnOpen.oncommand = function(){
var path = fsys.dlg.open("图片文件|*.jpg;*.png;*.bmp;*.webp",,,winform);
if(!path) return;
if(currentMat) currentMat.release();
currentMat = cv2.imread(path); //读取图片
if(currentMat && !currentMat.empty()){
winform.infoOverlay.text = "";
updateDisplay(currentMat); //显示图片
winform.statusText.text = "已加载图片，点击右侧按钮开始检测";
winform.statusText.redraw();
winform.infoOverlay.redraw();
}
}
winform.btnDetect.oncommand = function(){
if(!currentMat) return;
winform.btnDetect.disabledText = {"✶";"✸";"✹";"✺";"✹";"✷"};
winform.setTimeout(function(){ //防止按钮多次点击
doDetection(); //检测人体
winform.btnDetect.disabledText = null;
}, 20);
}
winform.onClose = function(){
if(currentMat) currentMat.release();
if(globalNet) globalNet.release();
}
// ====== 启动 ======
winform.show();
winform.statusText.text = "正在初始化系统...";
winform.statusText.redraw();
prepareModel();
win.loopMessage();
超赞