Image Shadow Removal Based on Residual Neural Network

Abstract

The removal of image shadows has always been a challenging task that requires us to detect shadows and understand the surrounding scenes. The existing method of shadow detection and removal first locates the shadow area by shadow detection,and then uses some reconstruction algorithms to remove the shadows of the umbra and penumbra. However, detecting shadows is already a very rare task. Based on the traditional physical methods can be applied to a high quality image, and a method based on statistical characteristics must manually tag shadow. In this paper,we use a convolutional residual neural network to train the model. Using the residual neural network, we can prevent degradation due to the excessive number of network layers. The trained model can detect the shadow area by inputting the global image and combining the semantics of the picture. In these two aspects, good shadow area detection and positioning can be obtained, and image shadow removal can be realized.