微调PP-OCRv3文本识别模型，对图片竖着文字识别效果差，训练数据集需要特别处理吗？

[zxy2020]

python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec.yml -o \
    Global.pretrained_model=ch_PP-OCRv3_rec_train/student.pdparams \
    Global.save_model_dir=./output/rec_ppocr_v3_exp2 \
		Global.epoch_num=200 \
		Global.checkpoints=output/rec_ppocr_v3/best_accuracy.pdparams \
    Global.eval_batch_step="[0, 90]" \
    Optimizer.lr.name=Const \
    Optimizer.lr.learning_rate=0.0005 \
    Optimizer.lr.warmup_epoch=3 \
    Train.dataset.data_dir=./ \
    Train.dataset.label_file_list=[./train_data/rec/rec_gt_train.txt] \
    Eval.dataset.data_dir=./ \
    Eval.dataset.label_file_list=[./train_data/rec/rec_gt_val.txt]

[rizwan777]

yes !! we definitely need to consider alignment of text.

[zxy2020]

yes !! we definitely need to consider alignment of text.

thanks you reply! could you tell me the url of doc about alignment of text?
谢谢你的回复，请问哪里有关于文本对齐的文档或者说明？

[rizwan777]

actually i dont have that in hand.

but i have worked before in similar project that i used to preprocess the image (text).

Preprocessing:
Depending on the layout of the Chinese text (vertical or horizontal), you might need to perform additional preprocessing steps, such as rotating the image if the text is vertical. OpenCV provides functions to perform image rotations.

Rotate the image clockwise by 90 degrees

rotated_image = cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE)

[zxy2020]

thank you ! I find the code at infer process

    dst_img_height, dst_img_width = dst_img.shape[0:2]
    if dst_img_height * 1.0 / dst_img_width >= 1.5:
        dst_img = np.rot90(dst_img)
    return dst_img