Combined label matrix with the conditional generative adversarial network for secret image restoration
Jianzhong Yang, Xianquan Zhang, Chunqiang Yu, Guoxiang Li, Zhenjun Tang · Alexandria Engineering Journal · 2025
Because the noise in a corrupted secret image is very special, the existing denoising algorithms have difficulty directly restoring the corrupted secret image well. To improve the recovery effect, we propose a secret image restoration algorithm that combined the label matrix and the conditional generative adversarial network (CGAN). The trustable pixels are extracted from the corrupted secret image and formed the label matrix. The label matrix is treated as an additional condition and concatenated with the corrupted secret image fed to CGAN, including the convolutional block attention module (CBAM) and residual block, for guiding the neural network to generate an acceptable synthesized image. Then, we use the synthesized image as reference image. For one noise pixel in the corrupted secret image, the pixel value of the corresponding position in the synthetic image is used as a reference value, and the closest one to the reference value among the candidate values of the noise pixel is selected to replace the noise pixel. The experimental result shows our algorithm outperforms other algorithms both the objective metrics and visual quality.