Research On Image Super-resolution Reconstruction Algorithm Based On Deep Learning

Image resolution is an important index to measure the richness of image information.However,the resolution of the original image obtained directly in the real world is usually difficult to meet the demand of practical applications,so the image super-resolution algorithm comes into being.The image super-resolution reconstruction based on deep learning aims to improve the quality and visual effect of the reconstructed image through algorithm design.In addition,in order to reduce the memory occupation of the reconstruction algorithm model,it is of great significance to research the lightweight and practical super-resolution algorithm.In this paper,we design the image super-resolution reconstruction models based on convolutional neural networks adapted to their characteristics for visible images and infrared images respectively,which dedicates to enhance the image resolution and reduce the parameters of the reconstruction model and computational complexity of the algorithm.The main research contents include:1.To address the existing super-resolution algorithm models with huge parameters and the problem of difficulty in extracting hierarchical feature information,we propose an asymmetric dilation convolutional residual distillation network for visible images.Firstly,we design two lightweight blocks called asymmetric shallow convolution residual block and asymmetric dilation convolution residual block respectively,which can expand the receptive field of the network and reduce the number of model parameters.Moreover,the combination coefficient of information distillation blocks corresponding to these two lightweight blocks as the core extraction layer is investigated to maximize the extraction efficiency of hierarchical features.In addition,in order to strengthen the correlation between hierarchical feature information and ensure the effectiveness of feature extraction,we introduce the channel shuffle mechanism and multi-scale spatial attention mechanism at the end of the information distillation block.Experimental results show that compared with other mainstream methods,the proposed algorithm recovers higher quality images with fewer network parameters and achieves a better trade-off between network parameters and reconstruction performance.2.To address the difficulty in assembling a large number of high-quality infrared image sample sets,we adopt the available visible images combined with fewshot infrared images and propose a progressive compact distillation network based on transfer learning.The introduction of the parallel dilation convolution can increase the perceptual field of the network while minimizing grid effects to obtain rich detail texture information.For global structure design,we propose the cascade connection to preserve shallow low-frequency features and the difference calculation algorithm between two adjacent blocks to extract high-frequency features.Furthermore,we propose the bil-global connection to accelerate network convergence.In the training stage,the transfer learning strategy is adopted to take the visible images and the fewshot infrared images as the pretrained network and fine-tuning network training datasets,respectively,to achieve super-resolution reconstruction of infrared images.The experimental results show that the proposed algorithm significantly improves the generalization ability of infrared image reconstruction after adopting the transfer learning strategy and achieves good performance in both subjective visual effects and objective index evaluation.