Research On Image Super-resolution Reconstruction Based On Feature Fusion

Images with higher pixel density can provide more detailed information and conform to human visual experience,so now people are increasingly pursuing high-quality images.However,due to the limitations of hardware conditions and external environment,it is still impossible to capture higher pixel images in many application scenarios,and the proposed image super-resolution reconstruction technology can effectively improve the image quality.The goal of this technology is to reconstruct the captured or existing lowresolution(LR)image from the perspective of the algorithm into a larger and clearer highresolution(HR)image.Image super-resolution reconstruction algorithms not only improve the perceptual quality of images,but also facilitate the solution of other image processing tasks,such as image understanding and sentiment analysis,making this technology a hot research topic in computer vision.Currently,deep learning-based methods achieve better reconstruction performance in image super-resolution reconstruction tasks.However,due to the limited ability of convolutional neural networks to express features,it is difficult to fully reconstruct complex textures in low-resolution images,and the reconstructed images will also have curved lines and blurred edges.In response to the shortcomings of existing deep learningbased reconstruction algorithms,this paper proposes an image super-resolution reconstruction algorithm based on feature fusion.The main tasks are as follows:First,an image super-resolution reconstruction model based on enhanced residual feature fusion is proposed(Enhanced Residual Network,ERN).In order to improve the expressive ability of the network,an enhanced residual module is proposed,which is composed of deconvolution and convolution operations.The deconvolution operation can expand the image feature size and further extract more high-frequency details.The network uses a multi-level feature fusion structure to make full use of the input lowresolution image information to restore high-resolution images with more texture details.The entire network is optimized using the L2 loss function.After the training is completed,the test image is input to the trained model to obtain the reconstructed super-resolution image.The generator in the SRGAN model is replaced with the proposed ERN model,and an enhanced residual feature fusion algorithm based on generative confrontation is proposed(Enhanced residual network generative adversarial networks,ERNGAN),which further proves that the ERN model can effectively improve the quality of the reconstructed image.Secondly,an image super-resolution reconstruction model based on multi-scale feature fusion is proposed(Multi-scale feature fusion network,MFFN).In order to solve the problem that images recovered by a network with a single convolution still have blurring and other problems,the MFFN network adopts a multi-scale structure based on residual sets of residuals,enabling it to adaptively extract features at different scales,while feeding the features extracted at different sizes into a cavity convolution layer with different expansion rates,enabling the network to extract more detailed information contained in the input image.Since L1 loss is better at preserving the complex textures and local structures in images,the entire network is trained using the L1 loss function to update the parameters.The proposed algorithm was objectively evaluated and subjectively analysed on four standard datasets,Set5,Set14,BSD100 and Urban100.The results show that compared with models such as SRCNN,the ERN model,ERNGAN model and MFFN model in this paper have significantly improved in terms of objective indicators.In terms of visual effects,the reconstructed image can restore obvious texture features and sharper edge effects,which is closer to the real high-resolution image.