语义分割是计算机视觉领域的核心技术,通过对图像中的每个像素点进行分类,将图像分割成若干个具有特定语义类别的区域。近年来,卷积神经网络(CNN)不断取得突破性进展,利用深度学习方法处理语义分割问题展示出具大的潜力。首先从语义分割的定义出发,探讨了目前语义分割领域存在的挑战。在介绍CNN相关原理的基础上,详细对比了几种用于语义分割算法评测的数据集,并重点对近年来语义分割领域基于解码器、信息融合和循环神经网络的深度学习方法进行综述。最后进行总结和展望,阐述了未来语义分割领域在进一步丰富数据库场景、提高算法实时性和开展三维点云语义分割三方面的发展趋势。

Semantic segmentation, which classifies all pixels in an image and divides the image into several regions with specific semantic categories, is a key technology in the field of computer vision. In recent years, convolutional neural networks (CNNs) have been making breakthroughs and have demonstrated great potential in using deep learning to perform semantic segmentation. Herein, beginning with the definition of semantic segmentation, existing challenges in the field of semantic segmentation are discussed. Based on CNN principles, several datasets used for semantic segmentation algorithm evaluation are compared in detail, and recent deep learning methods based on decoders, information fusion, and recurrent neural networks in semantic segmentation are summarized. Finally, future development trends (e.g. enriching database scenes, improving real-time performance of algorithms, and researching the semantic segmentation) of three-dimensional point cloud data in semantic segmentation are summarized.