Analysis of Convolutional Neural Network Architectures for Image Denoising and Restoration

Convolutional Neural Networks have performed excellently in image processing applications such as Image Demising and Restoration. CNN's performance and accuracy depend on its design and architecture. This work analyzes delousing and image restoration using different CNN architectures. In this experiment, three popular CNN architectures, U-Net, Res Net, and Dense Net, are trained on different noise levels. In this section, we analyze each parameter for the proposed CNN models, i.e., filter size and depth, etc., to evaluate performance variations due to these hyper parameters. Our results indicate that with a longer network and large filter sizes, it may be possible to deny the images more accurately for restoration. This development set opened as we experimented with fine-tuning the other tasks (DE noising and restoration) using pretrained models. We provide empirical evidence that has shown that incorporating pretrained models can result in a substantial improvement in the performance of CNNs with a limited amount of data. This work represents a complete investigation of CNN architectures, such as DE noising and restoration, and explains why specific engineering choices are effective. These results can be useful in enhancing accuracy and performance for image processing tasks using CNN models.