Instance Segmentation Algorithm of Urban Street Scene Based on Data Augmentation and Feature Enhancement
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 urb...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Harbin University of Science and Technology Publications
2024-04-01
|
| Series: | Journal of Harbin University of Science and Technology |
| Subjects: | |
| Online Access: | https://hlgxb.hrbust.edu.cn/#/digest?ArticleID=2309 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | 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. |
|---|---|
| ISSN: | 1007-2683 |