Distortion function for distributed deep image compression

Distributed Deep Image Compression (DDIC) is an emerging technique that utilizes inter-image correlations within the decoder to achieve efficient transmission and storage. The choice of distortion function plays a crucial role in the rate-distortion performance of DDIC, yet existing work lacks in-depth research specifically focused on distortion functions for DDIC. In order to further enhance the rate-distortion performance of DDIC, we propose a novel hybrid distortion function that employs a weighted combination of structural-level and pixel-level distortions, aiming to enhance the preserved structural features of compressed images while considering pixel-level characteristics. During the training process, the robustness and flexibility of the model are improved by balancing the attention on the dual features. The proposed hybrid loss demonstrates superior performance compared to other approaches on the KITTI and Cityscapes datasets.