A Novel UAV Aerial Vehicle Detection Method Based on Attention Mechanism and Multi-scale Feature Cross Fusion

With the rapid development of artificial intelligence science, more and more researchers try to use deep learning to train neural networks and have achieved great success in object detection. Vehicle detection based on UAV image is a special field of object detection. Due to the low resolution of the vehicle object, complex background, and less image information, it is challenging to extract robust visual and spatial features from the depth network and accurately locate the object in complex scenes. In this paper, combining the characteristics of vehicles in aerial images, we design a novel feature pyramid network called channel-spatial attention fused feature pyramid network (CSF-FPN) with Faster R-CNN as the basic framework. In CSF-FPN, a hybrid attention mechanism and feature cross-fusion module are introduced, so that feature maps can be generated with enhanced spatial and channel interdependence to extract richer semantic information. After our CSF-FPN is integrated into the Faster R-CNN network, the detection performance of small objects is greatly improved. The experimental results based on the VEDIA Dataset showed that the proposed framework could effectively detect the vehicle in large scene azimuth. Compared with the existing advanced methods, mAP and F1-score are improved.