An Image Compression and Encryption Approach with Convolutional Layers, Two-Dimensional Sparse Recovery, and Chaotic Dynamics

This research presents a method to encrypt and compress images using two-dimensional sparse decomposition, chaotic systems, and convolutional layers. The original image is first encrypted via the convolutional layers in the proposed approach, inspired by deep neural networks. This step is called CLE, which stands for convolution layer-based encryption, as it helps to increase the security level of encryption. Then, the encrypted image is compressed by two orthogonal matrices created by the chaotic systems and singular value decomposition. After that, the compressed image is quantized and mapped. In the next step, the pixels of the quantized matrix are scrambled to reduce the correlation between neighboring pixels. Then, the XOR operation is applied for the final part of the encryption process. In the decryption process, the direct inverse operators cannot be used. Therefore, the convolutional layer-based sparse decomposition (CLSD) approach is proposed to recover the original image by half quadratic splitting approach. Due to HQS, the proposed decryption approach is converted to two repetitive steps. The simulation results and the security analyses demonstrate that the proposed image reconstruction method performs well for various compression ratios.