| --- |
| license: mit |
| metrics: |
| - accuracy 99.51% |
| tags: |
| - OrientationDetection |
| - TextOrientation |
| --- |
| |
| Updated CNN .keras model with input layer 256*256 RGB, and 1 output node ranging from 0 (no rotation) to 1 (180 degrees). |
| |
| Its accuracy on the dataset in the link below is 99.51%, and the AUC (Area Under Curve) metric is 0.9990. |
| https://drive.google.com/file/d/1WENDw-Ggwy32KIdMlMmuF6NcO0gH3mC8/view?usp=sharing |
| |
| It only recognizes 0 vs 180 degrees as there's another robust way to distinguish +- 90 from 0 and 180 degrees (another model under UTSC DSU). |
| Additionally, the dataset only contains upright images because the training program generates the 180-degrees counterparts on the fly. |
| As you will see, the model is only trained on text slices resulted from my proprietary preprocessing methods |
| (old version already on GitHub https://github.com/digitalutsc/ml_image_tool, but I will update it soon), |
| So directly feeding the page images to the model may not work well. |
| |
| I used MobileNetV2 with imagenet weights as the backbone, and the accuracy proves the method to be very successful. |
| |
|  |
| |
