EasyOCR not reading Japanese text

Python Help
1

For some reason unknown to me EasyOCR will or will not at times read clearly legible Japanese text at random and I need help getting to the bottom of why? Can anyone help?

combine_images
combine_images2484×724 381 KB

    img = cv2.imread(imgap)
    if img is None:
        print("Image not found at path")

    # Increase contrast
    contrast = cv2.convertScaleAbs(img, alpha=1.5, beta=1)

    # Greyscale
    gray = cv2.cvtColor(contrast, cv2.COLOR_BGR2GRAY)

    # Denoise the image with Gaussian Blur
    blurred = cv2.GaussianBlur(gray, (3, 3), 0)

    # Threshold the image (binary)
    _, threshold_img = cv2.threshold(blurred, 128, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

    # Calculate the percentage of black pixels
    import numpy as np
    black_pixels = np.sum(threshold_img == 0)
    total_pixels = threshold_img.size
    black_ratio = black_pixels / total_pixels

    # Invert the image if black pixels dominate
    if black_ratio > 0.75:
        processed_img = cv2.bitwise_not(threshold_img)
    else:
        processed_img = threshold_img

    # Perform OCR on the processed image
    reader = easyocr.Reader(['ja'], gpu=True)
    result = reader.readtext(processed_img, text_threshold=0.6, low_text=0.2)
    print(result)
2

I ran your code about 10 times and it detected the text every time. are you getting any warnings or other output during the detection process?

3

I am getting nothing at all. That’s the problem or at least not the text that is displayed in the image. Sometimes it outputs garbage.

Edit: upping the LOW_TEXT value has helped the issue somewhat

4

out of interest try running this on your CPU not your GPU. It’s possible the software downloaded a different variant of the model based on our GPUs.

It’s possible (though entirely a guess on my part) that the library downloaded a lower parameter count model for you.