Incorrect inference results

[as-stevens]

Hi All,

I am trying to read the images as text using the below code:

image_path = 'test_sample.jpg'
img = read_img_as_tensor(image_path).unsqueeze(0)
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
predictor = ocr_predictor('db_resnet50', 'crnn_vgg16_bn', pretrained=True).cuda(device)
img = img.cuda(device)
result = predictor(img)

page = result.pages[0]
        #print(page)
height, width = page.dimensions
page_media = (0, 0, width, height)
for b in page.blocks:
    for line in b.lines:
    print(line)

The inference results are not good:

• The words are getting concatenated - for e.g. Membership interest inferred as Membershiplnterest
• Wrong characters identified- for e.g. Membership interest inferred as Membershiplnterest -note i from interest changed to l
• Missing text - some block of text is not read at all

Please let me if I need to do any pre-processing on the image or any other steps so that my results are more reliable.

[felixdittrich92]

Hi @as-stevens 👋 ,

This is a known issue so i think it would be great if you can wrote this also into #330 and add one or two examples if possible ? 🤗

Do you have two GPUs on your device ? Otherwise you can remove all without ocr_predictor(...).cuda() and have you tried to read your image with img = DocumentFile.from_images([image_path]) ? 😄