Using Asymmetric Attention Mechanism Access Class Barring with Gated Recurrent Unit for an End-to-end Image Compression

The exponential expansion of the Internet has resulted in an overwhelming abundance of multimedia data. However, image and video compression technology faces challenges such as insufficient feature extraction, inaccurate bit rate estimation, and quantization errors. Within this study, we introduce optimization methods to address these issues. The AT-ACB module is introduced to enhance compact feature extraction using an Asymmetrical Convolution Block and hybrid attention module. Besides, the parallel conditional context module captures global relevance and reduces spatial redundancy. Then, the Tiki-Taka training method mitigates quantizer mismatch. This paper mainly researches the end-to-end image compression optimization scheme built upon deep learning. Finally, the experimental results demonstrate that the proposed optimization method achieves BD-Rate savings of 3.95%, 0.97%, and 3.18% on the Kodak dataset, resulting in a total BD-Rate saving of 6.89%.