Research And Application Of Image Super-resolution Algorithm Based On Attention And Feature Aggregation

Image super-resolution technology can effectively improve image quality,and has been widely used in video image compression transmission,medical image,satellite image,public security monitoring and other fields.In recent years,thanks to the achievements of convolutional neural networks and generative adversarial networks in image super-resolution reconstruction,the quality of image reconstruction has been greatly improved.However,due to the huge parameters brought by the increase of network depth and the low utilization rate of image feature information,there are some problems in the reconstructed image,such as blurred edge texture,lost details and slow training speed of the reconstructed model,which further affects the application of image reconstruction.In view of the above existing problems,this paper conducts in-depth research on single image super-resolution reconstruction methods.Based on the image super-resolution reconstruction network SRGAN,a lightweight single view super-resolution reconstruction method is proposed and applied to the image super-resolution reconstruction process.The main work of this paper includes:1、Aiming at the problems of texture blur and redundant model parameters in SRGAN generated images,a lightweight image super-resolution reconstruction algorithm AL-GAN based on fusion attention mechanism was proposed.Firstly,the algorithm extracts the shallow features of the image,and combines the spatial channel attention mechanism to improve the ability of the deep feature extraction module in the generator,enabling the network to adaptively acquire high-frequency image feature information,thereby improving the performance of the generated network.Secondly,the network layers of the discriminator module are compressed,and more efficient convolution methods and activation functions are adopted to accelerate model convergence and improve the quality of reconstructed images.Finally,the mainstream DIV2 K data set in the image super-resolution task is used as the training set and verification set.The experimental results show that the improved algorithm is superior to the classical algorithm Bicubic and the deep learning algorithm SRCNN and SRGAN in reconstruction effect.The improved model parameters are greatly reduced,and the texture details are restored more clearly,which verifies the effectiveness of the improved algorithm.2、Aiming at the problems of low utilization rate of image feature information and poor visual perception of generated images in AL-GAN algorithm,a lightweight image super-resolution reconstruction algorithm HL-GAN based on hierarchical feature aggregation is proposed.Firstly,using residual networks and attention mechanisms to jointly build a generator module,and the high-frequency feature information of the image is increased by strengthening the communication between feature information of different scales,so as to improve the reconstruction ability of the network.Secondly,the markov discriminator(Patch GAN)with fewer parameters is used to replace the original discriminator,which makes the network training faster and more stable.Finally,the DIV2 K dataset is used for validation,and the experimental results showed that the improved algorithm improves the clarity of the image while preserving real details,enhances the visual effect of the reconstructed image,and further reduces the parameter count of the model.3、Combined with the theory of image super-resolution reconstruction,AL-GAN reconstruction algorithm is adopted,a prototype system of image super-resolution reconstruction based on generative adversarial network(GAN)is designed and developed.The system function module includes four parts: user interaction,image loading,image processing and system reconstruction.The system is able to perform super-resolution reconstruction of the input blurred images using the AL-GAN algorithm model,thus further improving the accuracy of the reconstructed image.