Research On Shadow Removal Method Of Natural Scene Based On Deep Learning

Shadow removal of the complicated natural scene is a challenging task.Shadow is a common phenomenon when collecting images of natural scenes.Shadow in the image will lead to a series of problems such as insufficient color change and brightness or information loss.Shadow removal of natural images is important for computer vision tasks and other applications.Up to now,researchers at home and abroad have done a lot of research on shadow removal of natural scene images and put forward many effective solutions.However,the quality of shadow removal of complex backgrounds and when the image is severely degraded(with deep shadows)remains to be further improved.In recent years,deep learning technology has been successfully applied to large datasets.In this paper,deep learning technology is used to study the shadow removal of natural scene images.The research contents and innovations are as follows:(1)This paper proposed a shadow removal method based on attentive generation network.This method utilizes the structure of the generative adversarial network.The generator includes an attentive decomposition network and an autoencoder network.The attentive decomposition network is used to decompose the shadow region from the image.The autoencoder is used to generate a shadow-free image.Finally,the discriminator is used to compare the generated shadow-free image of the original no shadow image to determine the quality of the generated no shadow image.We innovatively design an image shadow decomposition layer based on attention.We utilize it to focus on the shadow position in the image as a prerequisite for the autoencoder network.Extensive experimental results demonstrate that proposed model is an advanced and effective image shadow removal method,which can effectively to remove the shadow in complex scenes.(2)This paper also proposed a shadow removal method based on image generation inpainting model.This method can be used as a reference to network training to replace the severely shadow area with appropriate simulation content.The model is a full convolutional neural network,which can handle shadows of any position or of different sizes in the test time.Experimental results show that the method presented in this paper is better than the existing method of the case of severe shadow removal.This method achieves realistic shadow removal and can provide realistic results of degraded images even in severe shadow cases,rather than accurately reconstructing the contents of no shadow areas.