---

写在前面的话

最近参加了个小比赛,大概意思就是用yolov5实现一个害虫识别.前期一切顺利模型训练,检测什么的都完成了,但遇到了一个问题,每次进行推理都得用人家封装好的py文件,而且里面很多功能是我们不需要的,于是我就在想能不能把yolo的检测模块写成一个函数,传入numpy的array即可得出检测结果坐标.在查阅了大量资料后终于完成,而且检测精度和之前无区别.利用晚上时间简单写了个界面demo

---

一、明确功能目标

大概参考下布局
主要有三个按钮和两个文本框作为图片载体.

二、UI设计

---

1.选择主界面

采用无功能栏窗口

---

到此我们的界面大体上完工了,下面介绍如何实现按钮功能,以及图片的显示

---

三、功能的实现

1.构建信号槽,为关闭按钮添加功能

---

创建一个类,继承界面文件里的内容.连接好各自触发的函数

2.剥离yolo模型载入代码,载入训练好的模型,封装成函数

另写一个py文件,载入模型.当然,在主文件载入也可以

from models.experimental import attempt_load
from utils.torch_utils import select_device
WEIGHTS = 'runs/train/exp8/weights/best.pt'

def get_model():
    # 选择设备
    device = select_device('')  # device 设备
    half = device.type != 'cpu'  # half precision only supported on CUDA
    model = attempt_load(WEIGHTS, map_location=device)

    stride = int(model.stride.max())  # model stride
    names = model.module.names if hasattr(model, 'module') else model.names  # get class names
    if half:
        model.half()  # to FP16

    return model, device, half, stride, names

3.剥离yolo推理代码,封装成函数可以被方便的调用

model, device, half, stride, names = get_model()
imgsz = check_img_size(IMGSZ, s=stride)  # check image size

@torch.no_grad()
#不进行梯度检测
        def pred_img(img0):
            # Padded resize
            img = letterbox(img0, imgsz, stride=stride, auto=True)[0]
            # Convert
            img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
            img = np.ascontiguousarray(img)
            model(torch.zeros(1, 3, *imgsz).to(device).type_as(next(model.parameters())))  # run once
            img = torch.from_numpy(img).to(device)
            img = img.half() if half else img.float()  # uint8 to fp16/32
            img = img / 255.0  # 0 - 255 to 0.0 - 1.0
            if len(img.shape) == 3:
                img = img[None]  # expand for batch dim
            pred = model(img, augment=False, visualize=False)[0]
            # NMS
            pred = non_max_suppression(pred, CONF_THRES, IOU_THRES, None, False, max_det=1000)
            det = pred[0]
            im0 = img0.copy()
            gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwh
            annotator = Annotator(im0, line_width=LINE_THICKNESS, example=str(names))
            xywh_list = []
            if len(det):
                det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
                for *xyxy, conf, cls in reversed(det):
                    c = int(cls)  # integer class
                    xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
                    xywh_list.append(xywh)
                    label = None if HIDE_LABELS else (names[c] if HIDE_CONF else f'{names[c]} {conf:.2f}')
                    annotator.box_label(xyxy, label, color=colors(c, True))
            im0 = annotator.result()
            print(xywh_list)
            return im0, xywh_list

4.使用Qlabel展示图片,并用QtWidgets.QApplication.processEvents()方法高速刷新

正常来说Qlabel只能接收本地保存的图片,但实际上可以用np.array转化到QImage格式

---

    monitor={
    'left':0,
    'top':0,
    'width':2000,
    'height':1300
}
    def openimage(self):
        while 1:
            img=sct.grab(monitor)
            img=np.array(img)
            shrink = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
            # cv 图片转换成 qt图片
            qt_img = QtGui.QImage(shrink.data, 
                              shrink.shape[1],  
                              shrink.shape[0],	
                              shrink.shape[1] * 3, 
                              QtGui.QImage.Format_RGB888)
            self.label.setPixmap(QtGui.QPixmap.fromImage(qt_img))
            QtWidgets.QApplication.processEvents()

解释一下代码:
使用mss库截图,转化为nparray格式,再用cv转化到rgb颜色格式.使用QtGui.QImage方法传入img,宽高,色彩等信息返回QImage格式.之后便可成功覆盖到Qlabel上.循环末尾记得加上QtWidgets.QApplication.processEvents()插入线程,才能实现高速刷新实测效率很高,仅有微小延迟.
界面+实时截图占用cpu30左右

5.实时显示原图+显示推理后结果

本来当时想用多线程解决.但后来相信觉得没必要,直接顺便赋值一下给label就可以了.

    def startyolo(self):
        @torch.no_grad()
        def pred_img(img0):
            # Padded resize
            img = letterbox(img0, imgsz, stride=stride, auto=True)[0]
            # Convert
            img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
            img = np.ascontiguousarray(img)
            model(torch.zeros(1, 3, *imgsz).to(device).type_as(next(model.parameters())))  # run once
            img = torch.from_numpy(img).to(device)
            img = img.half() if half else img.float()  # uint8 to fp16/32
            img = img / 255.0  # 0 - 255 to 0.0 - 1.0
            if len(img.shape) == 3:
                img = img[None]  # expand for batch dim
            pred = model(img, augment=False, visualize=False)[0]
            # NMS
            pred = non_max_suppression(pred, CONF_THRES, IOU_THRES, None, False, max_det=1000)
            det = pred[0]
            im0 = img0.copy()
            gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwh
            annotator = Annotator(im0, line_width=LINE_THICKNESS, example=str(names))
            xywh_list = []
            if len(det):
                det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
                for *xyxy, conf, cls in reversed(det):
                    c = int(cls)  # integer class
                    xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
                    xywh_list.append(xywh)
                    label = None if HIDE_LABELS else (names[c] if HIDE_CONF else f'{names[c]} {conf:.2f}')
                    annotator.box_label(xyxy, label, color=colors(c, True))
            im0 = annotator.result()
            print(xywh_list)
            return im0, xywh_list
        while 1:
            img = sct.grab(monitor=monitor)
            img = np.array(img)
            img = cv2.cvtColor(img, cv2.COLOR_BGRA2RGB)  # 将图片通道类型学转为BGR
            shrink, aims = pred_img(img)
            qt_img = QtGui.QImage(shrink.data, 
                              shrink.shape[1],  
                              shrink.shape[0],	
                              shrink.shape[1] * 3, 
                              QtGui.QImage.Format_RGB888)
            qt_img1 = QtGui.QImage(img.data, 
                    img.shape[1],  
                    img.shape[0],	
                    img.shape[1] * 3, 
                    QtGui.QImage.Format_RGB888)
            self.label.setPixmap(QtGui.QPixmap.fromImage(qt_img1))
            self.label_2.setPixmap(QtGui.QPixmap.fromImage(qt_img))
            QtWidgets.QApplication.processEvents()
            

运行截图:

实时推理cpu占用极高,并没有什么实际作用.

结语

这个项目到这里也就结束了,目前也还是有很多问题的.例如点击关闭窗口后仅仅是ui关闭.开启的信号槽还在继续运行,必须手动ctrl+c一下结束程序.这个小项目用了我一个下午时间,算是个粗制滥造的东西.但应该不会再改善了把,哈哈