Rate-controllable Learned Image Compression Using Channel Attention

Classical learned image compression (LIC) methods usually require training multiple models to achieve the best compression performances at different rates, which greatly increases their training and deployment cost. Though existing methods can realize rate variation by using channel scaling factors or transform of the Lagrange multiplier, they are not able to adaptively control the compression process with desired rates, which causes additional trial cost if we want to obtain results with given compression ratios. In this paper, we address this issue by employing channel attention modules that use the desired target bit-rate as side information to adjust the distributions of feature channels, and a new rate-distortion loss function that integrates the target bit-rate into the rate-distortion optimization framework is proposed to train the model to realize continuous rate control. Additionally, a two-stage training strategy is utilized to ensure that the network can adaptively adjust the bit-rates, at the same time achieving the best rate-distortion performance. Experimental results demonstrate that our method achieves effective rate control over a wide range of bit-per-pixels (BPPs).