Research On UAV Aerial Image Object Detection Algorithm Based On Multi-scale Feature Fusion

UAV is becoming increasingly vital in the military,emergency rescue,intelligent agriculture,smart city building,and everyday life as UAV technology improves.Therefore,the study of high-performance detection algorithms for UAV aerial images has become a hot topic in recent research.with the deep learning’s superior learning ability,object detection technology has improved in recent years.New issues have arisen when applying deep learning-based object detection algorithms to UAV aerial image detection.In contrast to ordinary natural scene images,most of the objects in UAV aerial images are small and densely distributed,the scale of the objects changes widely,and the detecting background is also more complex.These issues have a significant impact on actual detection accuracy,making detection more complex and challenging.In this paper,we analyze the detecting difficulties of UAV aerial images and conduct research based on the YOLOv3 algorithm.The specific innovations are as follows:(1)Aiming at the problem of complex object background in UAV aerial images detection,a feature fusion module based on attention mechanism is designed.The module adds attention mechanism to help high-level and low-level feature maps pay more attention to the importance of their own features in the channel and spatial dimensions,and thus to eliminate complex background interference.The high semantic information from the high level’s feature map is then used to guide the recovery of the low level’s feature map’s detail information,helping the high and low level’s features to be fused more effectively.(2)Aiming at the problem of small target size and wide scale variation in UAV aerial images,a feature enhancement module based on cascaded dilated convolution is designed,and an adaptive channel and spatial feature fusion technique is provided.To obtain multi-scale feature information in various receptive fields,dilated convolution with different dilated rates is used,and then cascade fusion is used to improve the semantic representation of small objects.Meanwhile,the conflicting and redundant information generated by the fusion of different scale feature maps is filtered at the channel and spatial levels to increase feature scale invariance and significantly improve the detection effect on multi-scale targets.(3)Aiming at the application problem of UAV aerial image object detection,this paper makes its own DTX UAV aerial image dataset(DTX-ai)based on the DTX water conservancy hub construction project,and applies this paper’s improved UAV aerial image object detection algorithm to the UAV intelligent inspection system to detect abnormal phenomena in the construction area and surrounding waters.In this paper,ablation experiments and comparison experiments are conducted on the public aerial image dataset(Vis Drone)for the improved algorithm.The results of the ablation experiments show that each enhanced module is feasible and effective,and the results of the comparison experiments show that the improved algorithm in this paper outperforms existing standard object detection algorithms.The training and testing results on a homemade aerial image dataset(DTX-ai),as well as the application effect in the DTX UAV intelligent inspection system,are good and perform well in engineering applications.