Transformer-CNN Interactive Learned Image Compression for Resource-Deficient Industrial Visual Information Communication

The application of image compression in industrial information systems has significantly enhanced data processing efficiency and emerged as a supportive technique for realizing Industry 4.0. The learned image compression (LIC) approach has rapidly surpassed classical image coding standards in terms of coding performance. Recently, the research on LIC methods has primarily focused on precise latent feature distributions through entropy model design. However, limited efforts have been dedicated to improve the transformation methods. This article aims to address this gap by proposing more expressive nonlinear transformation networks and incorporating a bit allocation module based on transformer-CNN interactive structure. The objective is to assign additional bits to essential features, thereby ensuring the production of high-quality decoded images, even under the conditions of resource-deficient visual communication. Extensive experimental results demonstrate that our method achieves the state-of-the-art (SOTA) compression performance at low bit rates, surpassing the existing LIC methods and the latest coding standard.