Research On Image Shadow Removal Algorithm Based On Deep Learning

Shadow is a common Optical phenomena in life.The existence of shadow in the image provides illumination information in different scenes.It can help people better understand the content of the scene in the image.At the same time,it will also affect the quality of the image,which will bring a lot of trouble to the related algorithms of various image processing later.Therefore,it is crucial to develop effective shadow removal algorithms.In the early stages,shadow removal tasks based on traditional methods have complex computational processes and poor removal effects.In recent years,the rise of deep learning has brought new ideas to image processing,and these new methods have gradually become mainstream algorithms in shadow removal tasks.This article mainly studies the use of supervised deep learning methods to recover unshaded images from shaded images without prior information.The specific content is as follows:(1)In the structure of encoder and decoder based on Convolutional neural network,an image shadow removal network combining near gradient algorithm is proposed.By inputting paired datasets,the model learns the mapping relationship between shaded images and real unshaded images,and generates shadow images end-to-end.To this end,a proximal gradient descent module is first proposed to filter redundant information from the initial input features.Then,gated convolution was used in the network to reduce the overfitting risk of the model and screen features of different importance in the image.Finally,a cross stage feature fusion mechanism was introduced in each layer of the network to improve the accuracy of image shadow removal by fusing feature information from different levels.(2)In order to solve the problem that the Convolutional neural network has insufficient ability to control the complex global structure information in the shadow removal task,a Shadow Uformer network is proposed to improve the image shadow removal effect.In the input stage of the image,the cross attention feature fusion module is used to adaptively fuse features between different layers,and the attention mechanism is used to adaptively learn the correlation of features,helping the network capture richer feature information.Secondly,the convolution layer in the Convolutional neural network is replaced by a local enhancement module based on the Transformer framework,and a multi head attention mechanism with non overlapping shift windows is used to capture more global dependencies.Finally,use a modem to process shadow images in different scenarios.