SSCNN: Shuffle Siamese convolutional neural network for image restoration with image degradation source identification
T M Sivanesan, Natarajan Vijayaraj · The Imaging Science Journal · 2025
This paper proposes the Shuffle Siamese Convolutional Neural Network (SSCNN) for identifying image degradation sources. Initially, a blur pixel map is identifiied employing PyramidNet. The resulting map and the input blurred image are then processed by SSCNN, which combines Siamese Convolutional Neural Network (SCNN) and ShuffleNet, to identify the degradation source. Then, deblurring is done by kernel estimation. Thereafter, noise pixel identification is performed using a Deep Kronecker Network (DKN). Noise removal is achieved through a statistical model. Then Image inpainting is conducted utilizing context-conditional generative adversarial networks (CC-GAN), which is tuned employing Jaya Waterwheel Plant Algorithm (JWWPA). Lastly, super-resolution for image restoration is performed by employing FC-GAN. The proposed SSCNN achieved promising results, with Structural Similarity Index Measure (SSIM) of 0.805, Universal Image QualityIndex (UQI) of 0.890, a PSNR of 40.596 dB, and a Second-Derivative like Measure of Enhancement (SDME) of 60.163%.