Image-Adaptive Context Modeling for Compression Based on Implicit Neural Representations

Le résumé fourni par la source

Over the past decades, numerous approaches have been proposed to compress images while maintaining high image quality and reducing computational cost. Traditional approaches such as JPEG rely on linear transformations, whereas more recent neural network based methods often referred to as learned image compression achieve high performance by training on large datasets. Implicit Neural Representations (INRs) have also emerged as a promising alternative. INRs overfit a compact neural network to a single image, enabling competitive compression performance with a lightweight decoder. Additionally, context models have been used to predict image features in order to reduce the overall bitrate. However, existing INR-based methods typically utilize a fixed context across all images, which may limit its adaptability to image-specific structures. In this work, we propose an adaptive context modeling method that constructs image-specific contexts to improve the rate-distortion performance. The proposed adaptive context is integrated into the INR-based compression model, and its effectiveness is evaluated in comparison with conventional methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.