Open jsxyhelu opened 3 years ago
Sorry for the delayed response. Thanks for your issue. I checked it and also found this problem. The DNN model is used to detect qrcode as the following illustration:
For the three fail QRcode images, I found that the QRcode is successfully cropped, however, the decoded result is not stable ( several pixel shifts make a failing decode) because the curved QRcode makes key point alignment difficult.
We are trying to fix the problem.
According to the opencv_test_wechat_qrcode results, if DNN is not used, 33 pass; if DNN is used, 30 pass. What is the reason for the difference here? So what is the value of using DNN and training these models?