Research On Image Super-Resolution Reconstruction Based On Lightweight Neural Network

As a primary source of information and a medium for us to perceive the world,images are ubiquitous in our daily life.Image processing technology has been applied to various areas,such as security monitoring,remote sensing imaging,and medical imaging.Limited resolution of imaging devices,complicated environmental factors and the loss of image transmission will cause the problem of insufficient image resolution.In order to recovery details in images and to obtain high-quality clear images,Super-Resolution(SR)reconstruction technology is introduced.However,without prior knowledge of original scenes and imaging processes,SR becomes an intractable ill-conditioned inverse problem of the imaging degradation.By establishing a mapping relationship between high-resolution scenes and degraded images from a large amount of data,deep-learning-based methods have advantages in providing SR solutions.In this work,two novel lightweight super-resolution reconstruction models are proposed,as improvements of the Enhanced Super-Resolution Generative Adversarial Network(ESRGAN)on both network structure and edge enhancement,which have good performances in SR reconstruction with preferable visual effects and efficiency.Since ESRGAN relies heavily on computing resources due to its complex structure,to design a lightweight network,an optimized SR neural network based on an encoder-decoder structure is proposed as an improvement of ESRGAN.First,the discriminative network in ESRGAN is removed to simplify the network structure and to restrain false textures in reconstructed image.Moreover,the number of Residual-in-Residual Dense Blocks(RRDB)is cut down to reduce network parameters and speed processing.Besides,based on the structure of RRDB,the idea of encoder-decoder is used for image down-sampling and up-sampling to further reduce computing load.In the end,to avoid feature loss caused by samplings,skip-connections are used for preserving original image features as much as possible.Further,a SR reconstruction method based on edge enhancement is proposed to preserve rich texture information and reconstruct more details.The deep feature extraction module and the discriminant model in the ESRGAN are removed.An edge detection module is introduced for enhancing edge information instead.Since the Laplace operator extracts finer image edgescompared with other edge extraction operators,it is used for edge detection to enhance texture information of input images.In addition,an objective function that combines content loss,perceptual loss,and edge loss is proposed to make the entire network learn more sufficiently and efficiently.Experimental results show that the methods proposed in this paper effectively reduce the amount of network parameters and can achieve improved performances benefitting from their lightweight network structures.