Abstract: For the challenges of dense vehicle distribution, occlusion, and large scale variation in the vehicle instance segmentation task on UAV aerial images, a instance segmentation method based on feature enhancement and calibration is proposed. Firstly, for features extracted from backbone, in order to make full use of the high-level semantic features and improve the prediction effect of vehicles at each scale, Multi-scale Semantic Enhancement Module (MSEM) is proposed, which is added to the highest level of the feature pyramid structure to fuses multi-scale features and suppresses redundant information by using spatial attention. Secondly, for fused features after feature pyramid, in order to fully exploit global context information to improve the prediction of small-scale vehicles and suppress the noise introduced by the occlusion problem in local features, Global-Local Feature Calibration Module (GLFCM) is proposed, which consists of Global Feature Calibration Module (GFCM) and Local Feature Calibration Module (LFCM). GFCM apply channel attention and self-attention to achieve sufficient information interaction and calibration between feature levels, constructs dependencies between global feature maps and fully exploits global contextual information; LFCM adpots the multi-head self-attention mechanism Transformer for each level’s feature. Transformer in LFCM is linearized, which greatly reduces computation in comparison with original one, and at the same time has a better suppression effect on the occlusion noise in local features. Experiments are conducted on UVSD dataset, the proposed method improves segmentation precision by 5.1% for one-stage network YOLACT, 2.9% for two-stage network Mask R-CNN, and 1.6% for multi-stage network Cascade Mask R-CNN, which reflects that the proposed method can significantly improve the instance segmentation precision for vehicle on UAV aerial images.

Key words: UAV, Computer Vision, Instance Segmentation, Feature Enhancement, Attention Mechanism