The Research Of Image Super-resolution Reconstruction Algorithm Based On Transformer

Image super-resolution reconstruction refers to the technology of restoring one or more low-resolution images in the same scene to high-resolution images by using image processing and machine learning methods.As a basic task in the field of machine vision,super-resolution reconstruction has a wide range of applications.However,most of the existing super-resolution reconstruction methods are based on convolutional neural network for feature extraction,which has the problems of low efficiency of model representation and insufficient interpretation of extraction process,which limits the performance of the model.In recent years,although transformer has solved the above problems and achieved good results in image classification and segmentation tasks,there is still room for improvement in the adaptability of super-resolution reconstruction tasks.Therefore,in order to improve the performance of image super-resolution reconstruction model,combined with transformer technology,this paper designs the following two algorithms:(1)Image super-resolution reconstruction algorithm based on improved transformerIn order to improve the learning ability of the model to global features,this paper applies transformer technology to the super-resolution reconstruction task,and improves its structure on the basis of adapting to the task.Specifically,the model uses transformer instead of convolutional neural network to extract deep feature information,and replaces the key matrix and value matrix originally generated by low-definition image features in the attention mechanism with the original image features,so that the model can learn a more standard end-to-end image reconstruction method in the training stage.In addition,in order to avoid the stretching of feature channels in the process of image reconstruction,this paper improves the traditional residual structure,removes the batch regularization layer in the fusion process of shallow and deep features,and strengthens the feature fitting ability of the model.Experiments show that the adaptability improvement of transformer mechanism in super-resolution reconstruction task can effectively improve the performance of the model on common public data sets.(2)Super resolution reconstruction algorithm of transformer image based on multiscaleIn order to solve the problems of single feature extraction mode and insufficient utilization of feature information between tasks with different magnification in the current transformer super-resolution reconstruction model,a multi task scale feature fusion module is designed in this paper.Specifically,based on the improvement of transformer’s attention mechanism and combined with the characteristics of super-resolution reconstruction task,the model cascades the image feature information of different scales step by step,so that the image reconstruction process with high magnification can obtain more comprehensive characterization information.In addition,in order to adapt to the transformer super-resolution reconstruction task with multi-scale module,a hybrid loss function is designed to evaluate the feature similarity of different scale spaces.Experiments show that the improved design of multi-scale structure in transformer superresolution reconstruction task can make the model achieve better results in image detail reconstruction.This paper also applies the designed algorithm to the target detection task in the complex urban scene,completes the resolution improvement and information reconstruction of the input image,and verifies the effectiveness and practicability of this algorithm from the application level.