Study On Image Super-Resolution Reconstruction Method Based On Convolutional Neural Network

The spatial resolution of image is an important indicator to measure the image quality.In the process of imaging,due to the physical limitations of the imaging device and various external factors,it is difficult to obtain the high-resolution images that meet actual needs.Image super-resolution(SR)reconstruction techniques can overcome the inherent limitations of the imaging system hardware and achieve the purpose of reconstructing low-resolution images to high-resolution images through image processing.Thanks to the wide application of convolutional neural network(CNN)in high level computer vision tasks,the CNN-based SR algorithm has become the mainstream in the field of image SR.According to the up-sampling method adopted by the SR algorithm and its different positions in the network,our paper systematically collates and classifies the current SR algorithms based on CNN,and further carries out work from two aspects,namely,the network structure design strategy of the algorithm and the image reconstruction task in complex degradation scenarios,including:(1)In terms of the network structure design strategy of the algorithm,due to the current SR algorithm's improvement in reconstruction performance relies too much on the deepening of the network layer or the increase in the number of features,it overall reduces the utilization of image features in the network and cannot significantly improve the image reconstruction quality.This paper proposes an image SR algorithm based on deep multi-level up-projection network(DMUN).DMUN realizes the fusion of multi-level low-resolution image features in the local part of the network through the local feature up-projection unit,which improves the utilization rate of the image feature in the overall network.The residual up-projection group uses a recursive method to control the parameter cost of the model.The residual activation block combines the original residual block with spatial-and-channel attention mechanism to improve the feature expression ability of the network and further avoid the repeated use of redundant features by the network.Experiments show that DMUN can achieve a better balance between the reconstruction performance of the algorithm and the model parameter amount.(2)In terms of the image reconstruction task in complex degradation scenarios,due to the limitations faced by the current CNN-based SR algorithms that only focus on the image reconstruction tasks in single degenerate scene,our paper conducts a research on the noisy image SR reconstruction tasks and proposes a noisy image SR algorithm based on conditional generative adversarial network(NSRGAN).In order to better handle the noise information in low-resolution images,NSRGAN introduces a new noise intensity estimation network on the basis of DMUN to provide auxiliary constraints for the SR reconstruction process of noise images.In addition,our paper considers two real noise model based on the camera imaging process and the generative adversarial network to model more realistic and complex image degradation scenarios.Experimental analysis under various noise models show that,compared with the existing SR algorithms,NSRGAN exhibits better robustness in the more complex task of noisy image SR reconstruction.