DDocE: Deep Document Enhancement with Multi-scale Feature Aggregation and Pixel-Wise Adjustments

Digitizing a document with a smartphone might be a difficult task when there are suboptimal environment lighting conditions. The presence of shadows and insufficient illumination may reduce the quality of the content, such as its readability, colors or other aesthetic aspect. In this work, we propose a lightweight neural network to enhance photographed document images using a feature extraction that aggregates multi-scale features and a pixel-wise adjustment refinement step. We also provide a comparison with different methods, including methods not originally proposed for document enhancement. We focused on aesthetics aspects of the images, for which we used traditional image quality assessment (IQA) metrics and others based on deep learning models of human quality perception of natural images. Our deep document enhancement (DDocE) method was able to lessen the negative effects of different artifacts, such as shadows and insufficient illumination, while also maintaining a good color consistency, resulting in a better final enhanced image than the ones obtained with other methods.