| Vision is the most important sense organ for human to obtain external information,and image is an important part of visual information.Image quality determines the quantity and accuracy of information in image.Therefore,image resolution has become an important parameter in image processing.Super-Resolution(SR)reconstruction is to reconstruct the corresponding high resolution(HR)image through one or more low resolution(LR)image.Traditional super-resolution algorithms often have some shortcomings such as high complexity,many restrictions and fixed image size,which lead to unsatisfactory super-resolution effect and low image quality.In this paper,an improved deep convolution neural network(IDCNN)algorithm is proposed,which achieves good results in image super-resolution reconstruction.The main contents of the article are as follows:1.Discusses the research background and significance of image super-resolution reconstruction,introduces different image super-resolution algorithms,including three categories: interpolation-based,reconstruction-based and learning-based methods,explains their respective algorithm principles,advantages and disadvantages,and focuses on bicubic interpolation in interpolation method,which is often used in the pretreatment of image problems,the existing image can be scaled to the target size.The evaluation criteria of image quality are introduced.Subjective evaluation and objective evaluation can comprehensively reflect the reconstruction results of an image and ensure the fairness of the experimental comparison.2.The convolution neural network is mainly studied.Understanding the core idea of network model,local perception,weight sharing and downsampling are adopted for feature extraction;the basic structure of network is the repeated connection of feature extraction layer and feature mapping layer,which refines image features by downsampling and convolution,and synthesizes output at the final full connection layer.The parameters of network training are updated by back propagation algorithm,commonly used gradient descent method;activation function,as an important parameter in feature extraction and mapping,will directly affect the final image quality;five loss functions are compared and introduced,and the choice of loss function is different for different experimental objects.3.Based on the framework of SRCNN,an improved deep convolution neural network model is proposed,and the improvement and innovation points are explained in detail.In the feature extraction module,the improved algorithm uses randomized rectified linear unit(RReLU)to avoid the gradient dispersion and over-compression phenomena which often occur in existing algorithms;in the non-linear mapping part,the deep mapping structure is adopted to increase the number of network layers to 24 layers,and residual learning is addedto avoid the degradation of deep network,at the same time,the network training speed is improved.Finally,the experiment shows that residual learning can effectively reduce the convergence time of the network and improve the reconstruction effect;RReLU activation function optimizes the negative half-axis interval of the function,avoids the phenomenon of information compression,and is superior to other activation functions in PSNR and SSIM indicators;Compared with Bicubic,ScSR and SRCNN,IDCNN achieves better reconstruction results in objective evaluation and subjective evaluation.PSNR and SSIM are generally higher than other three algorithms in Set-5 and Set-14 test sets.The image texture is clearer and the illumination information is richer. |
