Image Denoising Using Feature Map Based Convolutional Neural Networks

Abstract

Noise corruption is inexorable in image recognition and it may heavily agitate the optical features of the capturing image. Convolutional Neural Networks (CNNs) have acquired good contemplation in the research of denoising. They possess various approaches to removing noise from a degraded image and to reconstruct an image with high visual quality. Because of its stratified learning and self-operating feature extraction capabilities, CNN has become useful for Machine Vision tasks. Furthermore, these methods mostly train a specified model for every noise level and desire multiple methods for denoising images with various noise levels. Large feature sets result in producing a noise effect, which causes an over-fitting of the network. A new method of Feature Map Based Convolution Neural Network (FMBCNN) has been introduced for image denoising. The quantitative analysis is performed by calculating four metrics of MSE (Mean Square Error), PSNR (Peak Signal to Noise Ratio), SSIM (Structural Similarity Index), and Entropy. They put forward a building block for Convolutional Neural Networks that tends to increase channel interdependencies at nearly no cost of computation. The weights accompanying a kernel also called a mask, are tuned to choose features dynamically. The performance of classification, segmentation, and detection modules is heavily influenced by feature selection.