The Datasets generated by trdg。
You just need to generate the training and test datasets into two different directories.
The amount of training data for all models is 10000.
| Five Words with Noise | Single Word with Noise (trained with same Size 100x32) | Three Words with Background |
|---|---|---|
 |
 |
 |
| Test Accuracy: 92.5% | Test Accuracy: 93.5% | Test Accuracy: 40.4% |
| batch_size=16, epoch=30 | batch_size=32, epoch=30 | batch_size=16, epoch=30 |
crnn_str_simple-ipynb is a simple CRNN implementation that can be used with a single word dataset with good results.
crnn_str.ipynb can be used to identify text images of different lengths, and can also modify parameters to select single channel (grayscale image) or three channel (RGB color image) for training and testing.
The pretrained models can be downloaded by Google Drive or Baidu Netdisk .
The document description is as follows:
| crnn_str_multi_v.pth | crnn_str_background.pth | crnn_str_single_simple.pth |
|---|---|---|
Trained by crnn_str.ipynb using 5-word dataset. |
Trained bycrnn_str.ipynb using 3-word dataset with background. |
Trained bt crnn_str_simple.ipynb using single word dataset. |