关键词: 深度学习, 语义分割, 空洞卷积, 多尺度

Abstract: The development of computer hardware has greatly promoted the development of computer vision. Convolution neural network has made remarkable achievements in semantic segmentation. However, the stacking  of multiple convolutional layers inevitably result in the loss of detailed information in the boundary of objects. In order to preserve boundary information as far as possible and improve the accuracy of image segmentation, a multiscale atrous convolution neural network model is proposed. The proposed model utilizes multiscale pooling to adapt to different scale targets in images. Besides, atrous convolution layer is used to learn target features, thus the accuracy of detailed information is improved, better segmentation results are obtained. Experimental results on the ISPRS Vaihingen dataset show that the proposed multiscale atrous convolution neural network is effective for target boundary fitting.

Key words: deep learning, semantic segmentation, cavity convolution, multiscale