The ocr funtion of PaddleOCR has a "bin" or "inv" parameter to invert the image or convert to black and white, it helped improved the detection but what really fixed it for me was using the "slice" parameter and giving it the appropriate sliding window for the size of my images. See doc https://paddlepaddle.github.io/PaddleOCR/latest/en/ppocr/blog/slice.html