Deep semantic image compression via cooperative network pruning

Incorporating semantic analysis into image compression can significantly reduce the repetitive computation of fundamental semantic analysis in downstream applications such as semantic image retrieval. In this paper, we tackle the semantic image compression task, which embeds semantics in the compressed bitstream. An intuitive solution to this task is joint multi-task training, which generally results in the trade-off of one task to accommodate the other. We thus provide an alternative pilot solution: given a pair of pre-trained teacher networks that specialize in image compression and semantic inference respectively, we first fuse both models to acquire an ensemble model and then leverage cooperative network pruning and retraining to condense the knowledge. Various experiments on five benchmark datasets validate that the proposed method achieves on par and in many cases better performance than the teachers yet comes in a more compact size, and outperforms its multi-task learning and knowledge distillation counterparts.