Optimal image Denoising using patch-based convolutional neural network architecture

Traditional CNN uses fixed location which is irrelevant and they show the inability to capture edges and texture which causes the smoothness of artefacts, thus many details are lost. Hence, this research work designs and develops a novel CNN-backed architecture i.e. PCNN (Patch-based CNN). PCNN sets network depth based on patch size. Moreover, research work follows various steps, first patch similarity identification is carried out later it is given to the designed customized CNN for denoising, and then these patches are integrated to achieve the efficient denoised image. Performance evaluation of the proposed PCNN architecture was carried out on the KODAK dataset by comparing and considering different parameters like PSNR and SSIM. A further different signal level is used for deep evaluation. Comparative analysis with different state-of-art techniques including deep learning-based shows that PCNN simply outperforms the other model. Further analysis is carried out on three random images i.e. image 8, image 19, and image 20 are considered for PSNR and SSIM comparison and it is observed that in terms of PSNR, PCNN achieves 93.03%, 62.01% and 67.78% improvised over the existing model. Further, SSIM is considered for the same images and PCNN achieves 42.55%, 19.84%, and 15.63% improvisation over the existing model. Further, this research compares PSNR values over different signal levels of 2, 4, 6, 8, and 10 and achieves improvisation of 45.51%, 38.72%, 38.57%, 56.2%, and 35.06% respectively. Furthermore, Average PSNR is compared considering Red, Green, Blue, and RGB, PCNN achieves improvisation of 5.65%, 2.34%, 8.62%, and 4.87%.