Convolutional Neural Network for Image Restoration Using Spatial and Multilevel Edge Features

Abstract

Image noise removal techniques have become an important research area in image processing because images are prone to blurring due to noise interference in the creation and storage process, and low-quality images increase the difficulty of obtaining information. Currently, the main solutions for image denoising can be divided into traditional denoising algorithms and deep learning-based denoising algorithms, and traditional denoising algorithms are relatively low in computation and faster in speed. Deep learning-based denoising algorithms have high complexity, but have better ability to maintain and denoise the texture information of images. Most convolutional neural networks (CNN) experience gradient loss as layers are stacked, resulting in poor network performance. Due to this problem, the network damages the edges and details of the objects in the image. To solve this problem, this paper proposes an in-image noise cancellation dual-path network using spatial and edge properties based on CNN. The entire network consists of two sub-networks, and the first is called the Edge Detail Enhancement Sub-network. For this subnetwork, the edge predicted using Edge-Net is used with the network's input along with the image. During learning, objects in the image are more clearly restored using clear edge information. The features of the image and edge are extracted through the double path block. The second subnetwork is called the Global Feature Capture Sub-network. The network consists of three blocks. Blocks include extended convolution and general convolution to improve denoising performance. By combining feature information between shallow and deep layers over long paths, we increase the information to be learned. Various features are learned by extracting image features and sharing different results of each path using a dual-path network. Experiments in this paper use images with three levels of additive white Gaussian noise (AWGN) with noise levels of 15, 25, and 50 as input to the network. To evaluate the performance of the network, objective image quality evaluations such as peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM) are used. Experimental results show a higher number than the reference noise removal algorithm. In the subjective quality comparison, the reconstructed image results in a clearer preservation of edge and texture details, with the proposed network eliminating more noise.