Learned Low Bit-rate Image Compression with Adversarial Mechanism

Adversarial mechanism is introduced to learned image compression system in this paper. Our motivation is that the number of quantization levels is limited with the constraint of low bit-rate, resulting in severe distortion in details after reconstruction. The adversarial training manner enhances the ability of Decoder/Generator to enrich textures and details in the reconstructed image. Channel-spatial attention mechanism is used to refine the intermediate features implicitly to boost the representation power of CNNs. As for entropy model, we jointly take hyperpriors and autoregressive priors for accurate probability estimation. Moreover, an EDSR-like post-processing subnetwork is concatenated after Decoder for further quality enhancement. The proposed approach demonstrates competitive performance when evaluated with multi-scale structural similarity (MSSSIM) and favorably visual quality at low bit-rate.