Self-attention and Unet-Based Lossy Image Compression

This paper proposes a deep learning-based image lossy compression algorithm based on U-Net and self-attention (AttUnet), which significantly improves compression performance through two core innovations. Cross-layer connections are introduced into the symmetric structure of the encoder-decoder, directly transmitting the features of each layer of the encoder to the corresponding layer of the decoder, effectively preserving image detail information and addressing the issue of high compression distortion in traditional autoencoders (AE). The self-attention module is used to model global dependencies, and dynamic weight allocation is employed to emphasize key features, thereby enhancing the representational capability of latent features. Experiments on the Kodak dataset demonstrate that the algorithm significantly outperforms JPEG, JPEG2000, and mainstream deep learning-based compression models (convolutional autoencoder (CAE) and AE-generative adversarial network (AE-GAN)) in terms of peak signal-to-noise ratio (PSNR) metrics. Compared to JPEG2000, it reduces the BD rate by 18.9%, and compared to CAE, it reduces it by 5.2%. By combining Huffman entropy coding with uniform noise quantization simulation, the algorithm offers a new approach for image compression that balances global feature perception and detail preservation.