Image Shadow Detection And Removal Algorithm Based On Deep Learning

With the development of deep learning,artificial intelligence has achieved rapid development,which makes automatic driving possible.In the process of image acquisition,shadow is inevitable due to the existence of different light intensity,occlusion and other factors.In the process of automatic driving,the existence of shadow has both advantages and disadvantages,so it is necessary to carry out the research of shadow detection and removal algorithm.This paper realizes shadow detection based on ARFPN and shadow removal based on FFTGAN.The main contents and conclusions of the research are as follows:(1)ARFPN algorithm based on feature pyramid is proposed to solve the existing problems of shadow detection.In order to solve the problems of small area objects,light shadow missing detection,dark pixel misdetection and boundary blurring,guided attention,adaptive feature fusion and encoder refinement modules are introduced.Verification was carried out on the current mainstream data sets ISTD and SUB,and the results were analyzed quantitatively and qualitatively.The experimental results show that the proposed ARFPN algorithm is superior to the current mainstream shadow detection algorithm.Compared with the DSDNet algorithm,the BER index of ISTD data set decreases by 8%,and that of SBU data set decreases by 6%,which proves the effectiveness and superiority of the proposed method in shadow detection.(2)Aiming at the existing problems of shadow removal,FFTGAN algorithm based on generative adversarial network is proposed.In order to solve the problem that a single learning feature is difficult to focus on texture,color,brightness,detail and other features at the same time,and the restored image is difficult to keep consistent with the surrounding environment,fast Fourier residual block and channel pixel attention mechanism are introduced in the generator.The current mainstream data sets ISTD and SRD were verified,and the results were analyzed quantitatively and qualitatively.Experimental results show that the proposed FFTGAN is superior to current mainstream shadow removal algorithms.On ISTD data set,RMSE index decreases by 32%compared with Mask-Shadow GAN algorithm and 18% compared with STCGAN algorithm,which proves the effectiveness and superiority of the proposed method in shadow removal process.