Coupled Squeeze-and-Excitation Blocks Based CNN for Image Compression

Recent researches have shown that deep convolutional neural networks (CNN) have achieved promising results in the field of image compression. In this paper, we propose an end-to-end image compression framework based on effective attention modules. In the proposed method, two channel attention mechanisms are employed jointly. The first is the Squeeze-and-Excitation block (SEblock) in the encoder. The other is the novel inversed SEblock (ISEblock) placed in decoder. These blocks, named coupled SEblocks, are placed behind the convolutional layer in both encoder and decoder. By using SEblocks, the encoder learns the interdependencies between different channels and the feature maps can be better distributed after entropy coding. In decoder, the inversed SEblock is employed which adaptively learns the weights and divides weights between the channels to supplement information compressed from the encoder. The whole network is trained as a joint rate-distortion optimization by using a subset of the ImageNet dataset. We evaluate our method on public Kodak test set. At low bit rates, our approach outperforms the existing Balle's, JPEG, JPEG2000 and WebP on multi-scale structural similarity (MS-SSIM) and gets good visual qualities for all images at test set.