This repository uses crnn to recognize capcha and an average pooling is added so that the input height can change to 64.
- python3
- opencv-python
- torch
- torchvision
- numpy
python3 infer.py --model_path crnn_capcha.pth --imgs_dir imgsTest on python generated 10000 images
elapsed time: 25.085253953933716, accuracy: 0.9523| Ground truth | Prediction | Image |
|---|---|---|
| gPKuG | gPKug |  |
| izyP | izyP |  |
| fM7n | fM7n |  |
| txjA | txjA |  |
| LWCN | lWCN |  |
| fa0Z | fa0Z |  |
| PKqOE | PKqOE |  |
| AkkHB | AkkHB |  |
| owMj | owMj |  |
| rtmXI | rtmXI |  |
| Ox2wG | Ox2wG |  |
| sL82v | sl82v |  |
| ncOy3 | nc0y3 |  |
| PLjz | PLjz |  |
| mU0Na | mU0Na |  |
Now the recognize errors are just some characters confuse, like 'x' and 'X', 'p' and 'P', 'o' and 'O'