Research On Face Super-Resolution Reconstruction Algorithm Based On Deep Learning

In the real world,there are often some images with relatively low resolution due to various reasons.These low-resolution images cannot meet the requirements of people’s daily life and production for image clarity in the era of rapid economic development,so image super-resolution reconstruction technology will play a vital role.Image super-resolution reconstruction technology can reconstruct low-resolution images to generate high-resolution images with better quality,so it has important application value in scenes that require high image quality,such as face recognition.Although the performance of the general image super-resolution reconstruction algorithm applied to the reconstruction of face images is good,considering the particularity of face images,these methods may cause reconstruction of human face images with texture blur and other problems due to the low utilization of the rich feature information in face images.The main innovations of this article and the results achieved by the experiment are as follows:(1)First,by comparing and analyzing the possible improvement points of the DCSCN network,this paper makes improvements in the depth and loss function of the DCSCN network,and propose the I-DCSCN network.Through experiments,it is found that the image reconstructed by the I-DCSCN network basically reaches the highest value in PSNR and SSIM,and is slightly higher than VDSR and DCSCN,but significantly higher than other comparison methods.Compared with the bicubic interpolation method,when the scale factor is 4,the average PSNR and SSIM of the method on the three test sets increase by 3.412 d B and 0.038 respectively.(2)After improving the depth and loss function of the DCSCN network and proposing the I-DCSCN network,the self-attention mechanism is introduced into the I-DCSCN network,and the SA-DCSCN network is proposed,which consists of three parts: shallow Layer feature extraction module,deep feature extraction module and reconstruction module,in which self-attention machine is used for deep feature extraction module.A new up-sampling method,sub-pixel convolution method,is introduced in the reconstruction module.This method can reduce the impact of the transposed convolution method that requires zero padding on the up-sampled image,and can reduce the amount of calculation and quicken the construction speed.The SA-DCSCN network model and the classic face super-resolution reconstruction algorithm,the general-purpose deep learning-based super-resolution reconstruction algorithm and the I-DCSCN algorithm reconstructed high-resolution images were subjectively and objectively evaluated and found that SA-DCSCN is higher than other comparison methods in subjective and objective evaluation standards.Compared with the bicubic interpolation method,when the scale factor is 3,the average PSNR and SSIM of the method increase by 4.428 d B and 0.043 respectively,and the PSNR increase value of Test3 is 5.373 d B.(3)Since the low-resolution face image used for testing is obtained by smooth down-sampling of the high-resolution face image in a fixed manner,it does not consider the reasons for the formation of low-resolution face images in the real world.The SA-DCSCN network was tested on the real face data set,and it was found that the method has good performance in reconstructing the actual image.Compared with the bicubic interpolation algorithm,the performance is improved by 21.98%,and the performance of the benchmark model I-DCSCN is also improved by 6.46%.