Abstract:

At present, deep full convolutional networks have made considerable achievements in the field of image semantic segmentation. However, with the repeated down-sampling operation, the detail information of the feature map is greatly lost, which affects the segmentation accuracy. In this paper, it designs a deep full convolutional network for semantic segmentation. The network structure consists of encoder and decoder. Some dilated convolution layers are introduced at the end of the encoder to reduce the loss of detailed information. In each stage of the decoding process, the low-order semantic features of corresponding dimensions are integrated, and the global features are integrated at the end of decoder to improve the segmentation accuracy. Finally, it uses the data-enhanced CamVid dataset to train and test the network. And the test results reach 90.14% PA and 71.95% mIoU. Experimental results show that the output of the network performs better in terms of regional smoothing.

Key words: semantic segmentation, feature fusion, full convolutional neural network, dilated convolution

摘要：

目前，深度全卷积网络在图像语义分割领域已经取得了瞩目的成就，但特征图的细节信息在多次下采样过程中会大量损失，对分割精度造成影响。针对该问题设计了一个用于图像语义分割的深度全卷积网络。该网络采用“编码器—解码器”结构，在编码器后端引入空洞卷积以降低细节信息的损失，在解码过程中融合对应尺寸的低阶语义特征，并在解码器末端融入全局特征以提升模型的分割精度。使用数据增强后的CamVid数据集对网络进行训练和测试，测试结果达到了90.14%的平均像素精度与71.94%的平均交并比。实验结果表明，该网络能充分利用低阶特征与全局特征，有效提升分割性能，并在区域平滑方面有很好的表现。

关键词: 语义分割, 特征融合, 全卷积神经网络, 空洞卷积