Text-Aware Learned Image Compression: Preserving Text Readability with Mask-Guided Transform

In recent years, many task-oriented image compression methods have emerged. However, few have focused on maintaining and restoring text region information within image compression networks. Two key challenges arise when compressing scene text images: 1) guiding the compression network to prioritize text regions and 2) minimizing distortion in text regions during lossy image compression. To address these challenges, we propose a learned image compression method targeting text regions. First, we introduce entropy-regulated mask generation for adaptive text-background separation, ensuring seamless boundaries between text and background in reconstructed images. Second, the mask-guided transform (MGT) is designed to retain semantic information in text regions, reducing distortion during compression and ensuring better preservation of text features in the reconstructed images. Extensive experiments demonstrate significant improvements in preserving text readability for applications like traffic sign recognition and autonomous driving scene understanding, as validated across two benchmark datasets.