基于特征与数据增强的城市街景实例分割算法

Instance Segmentation Algorithm of Urban Street Scene Based on Data Augmentation and Feature Enhancement

城市街景分割是智能交通领域中一项关键的技术,对于城市街景环境中的客观因素例如遮挡、小目标等问题,提出一种基于特征增强与数据增强的城市街景实例分割算法DF-SOLO(data augmentation and feature en-hancement SOLO)。针对遮挡问题,通过非对称自编-解码器架构对城市街景图像进行数据增强,与传统方法相比处理后的图像更贴近真实的源数据分布。针对城市街景中的小目标分割问题,引入特征加权和特征融合的思想,特征加权模块在特征处理过程中能够根据特征的重要程度赋予不同的权值,提高对重要特征的利用率;特征融合模块从更细粒度的角度进行多尺度特征融合以解决尺度敏感问题,提高语义特征的描述性。通过在Cityscapes数据集上的实验表明,提出的实例分割算法在保证实时性的同时相较于单阶段SOLO算法和两阶段Mask R-CNN算法的mAP值上分别提升2.1%和2%,改善了对小目标和遮挡目标的分割效果。

Urban street scene segmentation is a key technology in the field of intelligent transportation.For the objective factors in the urban street scene environment such as occlusion,small objects,etc.,a DF-SOLO(Data Augmentation and Feature Enhancement SOLO)instance segmentation algorithm of urban street scene based on data augmentation and feature enhancement is proposed.Aiming at the occlusion problem,the urban street view image is enhanced by the asymmetric self-encoder-decoder architecture.Compared with the traditional method,the processed image is closer to the real source data distribution.Aiming at the problem of small target segmentation in urban street scenes,the idea of feature weighting and feature fusion is introduced.The feature weighting module can assign different weights according to the importance of the features in the feature processing process,so as to improve the utilization rate of important features;the feature fusion module Multi-scale feature fusion is performed from a finer-grained perspective to solve the scale-sensitive problem and improve the descriptiveness of semantic features.Experiments on the Cityscapes dataset show that the proposed instance segmentation algorithm can improve the mAP value by 2.1% and 2%respectively compared with the single-stage SOLO algorithm and the two-stage Mask R-CNN algorithm while ensuring real-time performance.Improved segmentation of small objects and occluded objects.