Image preprocessing for OCR on glossy, curved and unevenly illuminated packaging

Uncategorized
1

I’m trying to preprocess smartphone photographs of consumer-product packaging before sending them to PaddleOCR.

The challenging images contain combinations of specular highlights, curved surfaces, perspective distortion, wrinkles, uneven illumination and blur.

I’ve tried standard preprocessing methods such as CLAHE, adaptive thresholding, sharpening, denoising, morphology and perspective correction. However, some transformations improve one image while significantly degrading another.

I’m looking for advice on designing a preprocessing pipeline that adapts to the image rather than applying the same transformations to every image.

In particular, I’m interested in:

- specular reflection/highlight suppression

- illumination normalization

- blur/degradation assessment

- cylindrical/curved text rectification

- determining whether an image is suitable for OCR

- selecting preprocessing operations based on measurable image characteristics

Are there established computer-vision approaches or papers that would be appropriate for this problem?

I’m especially interested in solutions that can be implemented with OpenCV or other practical open-source libraries.

2

If it is available, Please share the test image.

3

The part about one transformation improving one image while degrading another is exactly the hardest part of this problem. Packaging OCR in production environments has been a big focus of my work, and adaptive pipelines that score image characteristics before choosing operations are the direction that actually holds up across varied real-world conditions. For specular highlights specifically, inpainting after highlight detection tends to outperform simple suppression when text sits under the glare. Are you capturing images in a controlled environment like a conveyor line, or purely from handheld smartphone photos in the field?