数据增广+手势识别[Paddlehub+PaddleX]

import paddlehub as hubimport cv2import numpy as npimport globimport osimport randomimport argparseclass segUtils():
    def __init__(self):
        super().__init__()
        self.module = hub.Module(name="deeplabv3p_xception65_humanseg")    def doseg(self, frame):
        res = self.module.segmentation(images=[frame], use_gpu=True)        return res[0]['data']def randomCrop(frame, h, w):
    bh, bw = frame.shape[:2]    if bw - w > 0 and bh - h > 0:
        randx = random.randint(0, bw - w)
        randy = random.randint(0, bh - h)        return frame[randy:randy + h,randx:randx + w]    else:        return cv2.resize(frame, (w, h))def main(args):

    backlist = glob.glob(os.path.join(args.backdir, "*.jpg"))
    handlist = glob.glob(os.path.join(args.handdir, "*.png")) + glob.glob(os.path.join(args.handdir, "*.jpg"))    print("back image: ", len(backlist))    print("hand image: ", len(handlist))
    save_dir = args.savedir    if not os.path.exists(save_dir):
        os.makedirs(save_dir)
        SU = segUtils()        for handpath in handlist:
            img = cv2.imread(handpath)
            basename = os.path.basename(handpath)
            filename, ext = os.path.splitext(basename)
            mask = SU.doseg(img)
            mask[mask <= 1] = 0
            mask[mask > 1] = 1
            
            mask = np.repeat(mask[:,:,np.newaxis], 3, axis=2)
            h,w = img.shape[:2]            for i in range(2):
                rback = cv2.imread(backlist[random.randint(0, len(backlist)-1)])
                crop = randomCrop(rback, h, w)
                res = mask * img + (1 - mask) * crop
                newname = filename + str(i) + ext
                cv2.imwrite(os.path.join(save_dir,newname), res.astype(np.uint8))if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--backdir", type=str, required=True)
    parser.add_argument("--handdir", type=str, required=True)
    parser.add_argument("--savedir", type=str, required=True)
    args = parser.parse_args()
    main(args)

训练

至于训练，当然是直接paddleX啦，直接用GUI版本的快速验证自己的想法啦，快速的选择分类模型，开始吧~ 选择的MobileNetV3_small_ssld这个模型，batchsize32，训练了30个epoch，太过简单，就不写流程了。         

测试

一些测试结果图，效果完全OK哈哈哈

本地开着摄像头视频做了一些测试，效果还不错，但是忘了保存视频了，之后补上。

# 验证 上传一张test.jpg的图片进行测试吧!python predict.py