Pixel Level Image Encryption Based on Semantic Segmentation

Nowadays, a large amount of image information is transmitting on the internet, and people are focusing more on image security. These years, encryption algorithm based on “scrambling-diffusion” architecture has made great achievement in security performance. Nevertheless, there is no need to encrypt the full image. Only part of it matters, which is defined as Region of Interested (ROI). In this paper, we propose a pixel-level image encryption method based on semantic segmentation. We use a PSPNet model trained on Cityscapes dataset to gain the coordinate of each pixels in ROI, then we encrypt these coordinates with a novel chaotic encryption system we proposed. In this system, we improved the scrambling-diffusion architecture. Scrambling and diffusion share the chaotic sequence iterated by the four-dimensional hyperchaotic Chen system, and generate the required key flows independently. Our encryption system acts well in the test, almost all the coordinates of ROI are covered. Cryptanalysis shows its superior security.