关键词: 图像语义分割, 双向交叉熵, 路径聚合结构, 多尺度预测, 深度学习

Abstract: The deep full convolutional neural network based on encoder-decoder structure has made significant progress in image semantic segmentation. However, the path of transferring low-level positioning information in the deep network to the high-level network is too long, which makes it difficult to use low-level positioning information in the decoder stage to restore the boundary structure of the object. Aiming at this problem, a path aggregation structure used in the decoder part of segmentation network is proposed. This structure shortens the propagation path of low-level information to high-level information in the segmentation network and provides multi-scale contextual semantic information, so that the segmentation network can produce more refined boundary segmentation results. Aiming at the pro- blem that the softmax cross-entropy loss function often used in semantic segmentation is insufficient to distinguish samples with similar appearance, this paper reforms the softmax cross-entropy loss function and proposes a bidirectional cross-entropy loss function. Combining the proposed path aggregation Atrous convolutional network with the new loss function method can obtain better results on the PASCAL VOC2012Aug data set, which increases the mIoU value from 78.77% to 80.44%.

Key words: semantic , image segmentation;bidirectional cross-entropy;path aggregation structure;multi-scale prediction;deep learning