Stereo Network for Blind Image Super-Resolution
Guangyi Ji, Xiao Guang Hu · 2024
Single image super-resolution method using neural networks has achieved remarkable strides. However, most existing works rely on the architecture of Convolutional Neural Networks (CNNs) with shared kernel and the increasing of vertical depth, as result, super-image would loss high-frequency information. In every layer of human retina exists huge of neurons and some neurons in the same layer connect each other by horizontal neurons. Inspired by the neural network of human retina, a Stereo Neural Network (SterNet) is designed for blind image super-resolution. As the basic block of SterNet, Dynamic Filter Block (DFB) performs through unshared kernel, hence SterNet is able to easily obtain more spatial features and high-frequency semantic information. To expand the network width, one DFB with unshared kernels and one RRDB with shared kernels connect in parallel to construct Dynamic Filter and Rense Residual Blocks (DFDRB) and two DFDRBs are in parallel to form a Stereo feature extraction Block (SterB). At last, several SterBs are in series into a SterNet. Extensive experiments on several benchmarks show the effectiveness of the proposed method. This indicates that simulating the structure and operation of real neural networks is beneficial for improving vision application system.