基于多层级上下文信息的图像语义分割

Image Semantic Segmentation Based on Hierarchical Context Information

随着深度学习和卷积神经网络的应用,图像语义分割的性能得到了大幅度提升。但当前图像语义分割算法在语义信息利用率及语义类别区分度方面仍有欠缺。为了进一步提升语义分割算法的性能,提出多层级的上下文信息机制,使用多层级特征对长距离的依赖关系信息和局部性较强的上下文信息进行提取,增强卷积神经网络特征的信息丰富度与类别区分度。所提多层级上下文信息机制在典型街道场景数据集Cityscapes验证集上的分割精度达77.2%,实验证明了所提方法的有效性。

The wide use of deep learning and convolutional neural networks in recent years has been one of the main reasons for performance improvement in image semantic segmentation.However,the current image semantic segmentation algorithms have certain drawbacks.For example,the semantic information is not fully used,and the discrimination between different semantic categories is not large enough.Therefore,we propose a hierarchical context information mechanism to achieve better semantic segmentation performance.The long-range dependency information and local context information(extracted from the hierarchical features)are conducive to enriching information and discriminating among different types of semantic categories.Our experiments demonstrate the effectiveness of the proposed method.The proposed method achieves a segmentation accuracy of 77.2%on Cityscapes val dataset.