Abstract
Recent works on learned image compression (LIC) based on convolutional neural networks (CNNs) have achieved great improvement with superior rate-distortion performance. However, the robustness of LIC has received little investigation. In this paper, we proposes a complex-valued learned image compression model based on complex-valued convolutional neural networks (CVCNNs) to enhance its robustness. Firstly, we design a complex-valued neural image compression framework, which realizes compression with complex-valued feature maps. Secondly, we build a module named modSigmoid to implement a complex-valued nonlinear transform and a split-complex entropy model to compress complex-valued latent. The experiment results show that the proposed model performs comparable compression performance with a large parameter drop. Moreover, we adopt the adversarial attack method to examine robustness, and the proposed model shows better robustness to adversarial input compared with its real-valued counterpart.