Hi there
I want to run text recognise model in axon. I already tried EasyOCR and PaddleOCR. but EasyOCR is less accurate and PaddleOCR is slow and running over CPU. any other suggestion or example for OCR model which can be optimise for axon
I see this is the problem with OCR, since the text recognition step is being run for each detected line that further divided into multiple blocks, resulting in very slow FPS. I would recommend sticking with paddle OCR, as it provides better accuracy.
Are you working with images that contain large amounts of text, or is it just for some particular use case?
I need for reading cropped vehicle number plate image.
You can use paddleOCR’s latest english recognition model for recognising the text and for detection you can use LPRNet for detecting number plate, which would be efficient.
For this, you should export both models to ONNX first and then to rknn model for running on NPU of axon.
I hope you know, PaddleOCR or any other OCR usually provides 2 models, one for detection and creating bounding boxes around the texts present in the image and 2nd for recognising the text. Generally detection model is larger than recognition model. Also in your case you don’t need to detect all texts present in the image, but only the number plate, so a model to only detect that would be smaller and more accurate and for recognition PP-OCRv5 would be quite good and efficient.
1 Like