search

YukSing12 / Steel-Billet-Character-Recognition

3 stars 0 forks source link

About how to train #1

Open xfby2016 opened 2 years ago

xfby2016 commented 2 years ago

Excellent work, thanks to open source！ I have noticed that the original image size is very large and the characters are in a small area, so is the training image directly labeled with the original image?Need for cutting and positioning?

YukSing12 commented 2 years ago

This is a two-stage algorithm. It first detects character areas. Second, it recognizes the character from the detected area. As a result, detection model is trained by the original image and the recognition model is trained by the cropped image. More details can referenced from https://github.com/PaddlePaddle/PaddleOCR

xfby2016 commented 2 years ago

My understanding the detection model is a target detection network? Are you training with the original image?For example, 2448 * 2048? Is this require a very large gpu memory to complete the training process ?

xfby2016 commented 2 years ago

When I run the project，I get the following error： ppocr\lib\site-packages\paddleocr\ppocr\postprocess\rec_postprocess.py", line 66, in
for text_id in text_index[batch_idx][selection] IndexError: list index out of range any suggestions？Thank you！

YukSing12 commented 2 years ago

Maybe your recoginition output is empty. You can have a check. Also you can use the newest version of PPOCR.