Research On Image Shadow Removal Algorithm Based On Generative Adversarial Networks

In recent years,with the development of society,computer vision,one of the three major areas of current artificial intelligence research,has been rapidly developed,however,because the presence of shadows can have a certain impact on the robustness of computer vision tasks,detecting and removing shadows from natural images has become an indispensable and fundamental task in computer vision and digital image processing tasks.In this thesis,the following work is accomplished to address the problem of shadow detection and removal from single images:Firstly,to address the problems of missed detection and false detection in image shadow detection by ST-CGAN,an improved ST-CGAN based image shadow detection algorithm is proposed,which introduces the CBAM attention mechanism into the network,enabling the network model to learn and adjust the weights of each channel and pixel adaptively,allowing the network to focus more on key channel and spatial features,thus improving its model performance and accuracy.Experiments show that the proposed algorithm in this chapter achieves better performance and significantly improved visual effects on both SBU and UCF common benchmark datasets compared with the comparison of other three algorithms.Secondly,a multi-scale attentional residual network based on generative adversarial network is proposed as an image shadow removal algorithm for shadow removal in complex environments with color and texture distortion in shadow regions and incomplete removal of semi-shadow regions.First of all,feature extraction is performed using different size convolutional kernels to obtain feature information at different scales,and then the feature information of each channel is calibrated by SE-Net(Squeeze and extraction Network)to enhance the important feature information and weaken the non-important feature information.Numerical experiments show that the proposed algorithm can accurately recover the details of color,texture and shadow edges of shadow regions in complex shadow scenes,and the root mean squared error(RMSE)of shadow regions can reach The Root Mean Squared Error(RMSE)of the shadow region can reach 8.4,the Structural Similarity(SSIM)can reach 97.2%,and the Peak Signal to Noise Ratio(PSNR)can reach 33.97 d B.