Optimized Image Restoration Based On Residual Cascade Convolution Neural Networks

Abstract

Images, while acquiring, or by passing over analog media, suffers degradation, which may directly affect the quality of scenes taken. Despite numerous image restoration methods proposed, effective image restoration is a further challenging problem. In recent years, the wide application of Convolution Neural Networks (CNN) facilitates more effective image restoration methods. In this work, a novel Residual Cascade Convolution Neural Network (R-CCNN) is proposed for high quality image restoration. The proposed method considers a single trained model which can be considered as a filter for blind image de noising capable of removing all types of noises. The method has the advantage of being computationally less complex. The previous proposed neural network models for restoration requires more number of parameters which makes it relatively more complex. The proposed R-CCNN method assigns an optimal value for these parameters and proposes a method with minimal regularization and hence we need only a less number of parameters. The optimal values for these parameters are found using a global optimization algorithm, Flower Pollination Algorithm (FPA). Results obtained shows that the proposed model is more advantageous than other state-of-the-art restoration methods by checking peak signal-to-noise-ratio (PSNR) and structure similarity index metrics (SSIM) values.