Research On Image Shadow Detect And Removal Based On Conditional Generative Adversarial Network

In natural images,the presence of shadows can provide us with information about the scene and lighting conditions,as well as help us understand the scene in the image,but at the same time it makes image processing technically difficult.In the field of image processing,shadows are usually removed,and their removal is preceded by accurate detection and localization.With the development of deep learning,a new path has been opened for the study of image shadows.However,there are still many research bottlenecks in this field,for example,in shadow detection,shadow boundaries are not obvious,details are handled roughly,etc.;in shadow removal,the recovered shadow-free images have artifacts,lighting conditions change,etc.Therefore,an effective shadow detection and removal method is urgently needed to solve the above problems.Recent studies based on deep learning are only single for shadow detection or shadow removal,and do not consider the correlation between the two tasks,so to solve this problem this paper proposes a Multi-layer Conditional Generation Adversarial Network(ML-CGAN)for shadow detection,ML-CGAN)model for shadow detection and removal,which uses a combination of two CGAN networks,and the studies in this paper are all based on this network framework,as follows.(1)In order to emphasize the features of the shadow target and ignore the information that is not useful for the target region,the generator in this chapter introduces the attention module under the U-Net network base framework.The attention module proposed in this chapter extracts information from deep and shallow features at different scales before fusing them,uses the semantic information contained in the shallow feature map to guide the deep feature map to select important shadow location information,and adds the attention module to the jump connection before sampling on the network.Finally,compared with other algorithms,the experimental results show that the network outperforms other methods for both shadow detection and removal tasks.In shadow detection,the balance error ratio is reduced by 5.4% compared to the suboptimal algorithm;in shadow removal,the structural similarity reaches 0.9654.(2)In order to solve the problems of deep network depth and easy overfitting in the ML-CGAN model based on the attention mechanism in Chapter 3,this chapter simplifies the design of the generator structure based on the ML-CGAN framework and proposes a multi-scale cavity convolution module for extracting feature information at different scales,and the improved generator mainly includes: encoder module,multi-scale cavity convolution module,decoder module.In order to avoid the gradient disappearance,the encoder module adds the residual connection.The experimental results show that for the proposed network,the balance error ratio decreases by 16.9%in shadow detection compared with the method in Chapter 3;while in the shadow removal result,the peak signal-to-noise ratio of the recovered shadow-free image is28.1153,which is improved by 2.4%.Finally,the effectiveness of the residual structure and the multi-scale cavity convolution module are verified.