Image compression based on octave convolution and semantic segmentation

摘要

Lossy image compression based on deep learning usually contains stacking convolutional layers, pooling layers, and nonlinear functions. However, the feature map is obtained by the convolutional layer, which has a lot of redundancy, so we use octave convolution instead of vanilla convolution to improve compression efficiency. The feature map can be divided into high-frequency and low-frequency information. We use octave convolution to design an automatic codec to decompose the feature map into high-frequency and low-frequency information, which effectively improves the quality of the generated image. First, the semantic segmentation map of the input image is obtained by pre-training SegNet. The ComNet uses the original image and the semantic segmentation map to generate a low-dimensional representation, and the GenNet network utilizes the low-dimensional representation and the semantic segmentation map to estimate images. Then, the residuals between the reconstructed image and the original image are encoded. Finally, the reconstructed image and the decoded residual image are used to obtain the final high-quality reconstruction. Experimental results show that our method outperforms the existing image coding standards in terms of PSNR and MS-SSIM at different bit rates, and the reconstruction of images with complex textures and semantics has more obvious advantages.