Instance segmentation for vehicle in UAV aerial images based on feature enhancement and calibration

To overcome the challenges of dense vehicle distribution，occlusion，and large scale variation in instance segmentation of the vehicle in UAV aerial images，an instance segmentation method is proposed based on feature en⁃ hancement and calibration. Firstly，to make full use of the high-level semantic features extracted from the backbone and improve the prediction effect of vehicles of different sizes，a Multi-scale Semantic Enhancement Module（MSEM）is proposed，which is added to the highest level of the feature pyramid structure to fuse multi-scale features by densely connecting multiple dilated convolution blocks with different dilated rates and suppresses redundant information by us⁃ ing spatial attention，this module alleviates the loss of high-level semantic features and improves the multi-scale adapt⁃ ability of features. Secondly，to fully exploit the global context information of the features fused with the feature pyra⁃ mid to improve the prediction of small-scale vehicles and suppress the noise introduced by occlusion of local features，a Global-Local Feature Calibration Module（GLFCM）is proposed，which consists of a Global Feature Calibration Mod⁃ ule（GFCM）and a Local Feature Calibration Module（LFCM）. The GFCM applies 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. The LFCM adopts the multi-head self-attention mechanism Transformer for the features at each level. The Transformer in the LFCM is linearized，which greatly re⁃ duces computation in comparison with the original one，and at the same time has a better suppression effect on the occlusion noise in local features. Experiments are conducted on the UVSD dataset. It is found that the proposed method improves the segmentation precision by 5. 1% for the one-stage network YOLACT，2. 9% for the two-stage network Mask R-CNN，and 1. 6% for the multi-stage network Cascade Mask R-CNN，which shows that the proposed method can significantly improve the instance segmentation precision for vehicles in UAV aerial images. © 2023 AAAS Press of Chinese Society of Aeronautics and Astronautics. All rights reserved.