Research On Unsupervised Image Super-Resolution Based On Attention Mechanism

As we enter the era of high definition video images,people’s demand for image clarity is increasing.Using technology to enhance image resolution is obviously a more economical and feasible choice than replacing high-definition image acquisition and display equipment.Therefore,image super-resolution technology has become a research hotspot.Currently,mainstream image super-resolution deep models simulate image degradation by learning the mapping relationship between low and high resolution images.However,these supervised image super-resolution models have the problem of weak generalization when dealing with complex real-world image super-resolution tasks.In addition,deep networks enhance feature learning by stacking convolution layers,but too deep networks will generate too many parameters,consuming a lot of computing resources.Also,limited by the local characteristics of convolution operations,they lack global correlation learning,leading to local distortion phenomena in reconstructed images.To address these issues,this paper studies the unsupervised image super-resolution reconstruction based on attention mechanism,and the main work is as follows:Firstly,an unsupervised image super-resolution reconstruction method based on pixel attention is proposed.The model uses pixel attention to weight the relationship between pixels,restores high-frequency details from low-frequency images,and reduces artifacts and jagged effects.At the same time,a segmented learning image degradation mechanism that integrates multiple degradation factors is designed,and the model is trained unsupervisedly with non-paired training images,which is validated on a benchmark test set.Compared with the baseline model,the PSNR is improved by an average of 0.79 d B,and the SSIM is improved by an average of 0.1,proving that pixel attention effectively improves the quality of reconstructed images.Compared with mainstream models,the PSNR and SSIM evaluation indicators show that this chapter model is superior to the comparisSecondly,an unsupervised image super-resolution reconstruction method based on multi-headed self-attention is proposed.Based on the effectiveness of the segmented learning image degradation mechanism,a real-world image degradation model with more complex degradation factors and more random degradation methods is further constructed.The model requires the reconstruction network to have stronger robustness,so this paper constructs a real-world image super-resolution reconstruction model based on multi-head self-attention mechanism,which constitutes a complete unsupervised real-world image super-resolution reconstruction method with the real-world image degradation model.Multi-head self-attention provides the network with powerful modeling capabilities,restores high-frequency details by modeling long-distance dependencies,and significantly improves the visual effects of reconstructed images.A comparative experiment was conducted on a classic super-resolution test set and a real-world image dataset,and the proposed model outperformed current mainstream unsupervised real-world image superresolution reconstruction models in terms of reference evaluation indicators PSNR,SSIM,and non-reference evaluation indicator NIQE,with an average improvement of 1.75 d B for PSNR,0.047 for SSIM,and a decrease of 2.47 for NIQE.Thirdly,an unsupervised image super-resolution reconstruction method based on fusion attention is proposed.Based on the existing real-world image degradation model and spatial attention mechanism research,the feature extraction fusion attention group is designed by combining channel attention.The fusion attention group uses residual connections to merge spatial and channel attention features,expands the receptive field,and further constructs an image super-resolution residual convolution network with scalability.Attention visualization results show that the network’s attention range to image feature-rich areas is expanded by 17-23% after passing through the fusion attention group.Extensive experiments were conducted on a benchmark test set and a real-world image test set,and the proposed method outperformed current mainstream models in terms of both reference evaluation indicators and non-reference evaluation indicators,with an average improvement of 1.63 d B for PSNR,0.045 for SSIM,and a decrease of2.93 for NIQE.The experimental results demonstrate that the proposed method in this chapter can achieve better performance than existing methods in terms of natural image quality evaluation.