MULTI-SCALE RESIDUAL CONVOLUTIONAL NEURAL NETWORK FOR SHADOW DETECTION IN HIGH RESOLUTION REMOTE SENSING IMAGES

摘要

The presence of shadows degrades the image quality and reduces the interpretation accuracy. With the improvement of the spatial resolution of the remote sensing images, the shadow effect has become more and more obvious. Therefore shadow detection is an essential step for remote sensing image analysis. Convolutional neural networks have recently demonstrated outstanding performance for image processing, but few works have been done for shadow detection in remote sensing images with this technology. In this paper, we propose a novel multi-scale residual convolutional neural network for shadow detection (MSR-SDCN) in high resolution remote sensing images. The core of the proposed network is composed of a multi-scale encoder network and a corresponding decoder network with residual structure. The encoder network is used to extract the image features and downsample the input feature maps gradually, while the decoder network learns to upsample its input feature maps and generates the shadow detection result which has the same size as the input image. The residual structure is used in two aspects, i.e., internal same scale connection and encoder-decoder connection. The training data is from UCF database and data augmentation is used to increase the amount of the training data with random downscale, flipping and rotation. The overall accuracy of the proposed method on the test images is above 90%. The shadow detection results from the proposed method are compared with the three shadow detection algorithms in visual analysis and quantitative assessment. It shows that the proposed deep learning approach can achieve more accurate result than the traditional methods.