Adaptive Clustering Object Detection Method For UAV Images Based On Deep Learning

With the integration and development of UAV technology and computer vision technology,object detection for UAV aerial images has become a research hotspot,and more and more detection algorithms with high detection accuracy and robustness have emerged.UAV aerial images have the characteristics of complex backgrounds,changeable scenes and a large number of small objects,which make the detection of UAV aerial images a difficult problem.In view of the above problems,this paper starts from two aspects: enhancing the UAV aerial image data set and improving the detection network structure,and is committed to improving the accuracy of object detection.By analyzing the characteristics of UAV aerial images,this paper proposes an adaptive clustering object detection method for UAV aerial images.The method is divided into three sub-networks: first,the adaptive clustering sub-network,In view of the characteristics of small object aggregation in UAV aerial images,this paper proposes an adaptive clustering network for extracting potential small object aggregation areas in UAV aerial images.The network outputs the segmentation of several small object aggregation areas.picture.Second,the segmentation and filling sub-network,evenly divides the pictures with too large size output by the adaptive clustering network,fills the pictures with unbalanced length and width ratio output by the adaptive clustering network,and performs the object area with too small size.Padding;the purpose of this method is to keep the size of the image within a reasonable range required by the detection network.Third,the detection sub-network,which is divided into a local detection network and a global detection network;the small objects in the pictures output by the segmentation and filling sub-network are aggregated,and the local detection network introduces an attention mechanism,uses NMS with variable thresholds,and uses sample balance.Strategy training and other methods effectively improve the detection accuracy of objects in the aggregation area;when the adaptive clustering network segments the image,it will truncate some large and medium objects,and the global detection network will truncate these truncated objects and the local detection network.The object is detected,and the result is a supplement to the results of the local detection network;the results of the local detection network and the global detection network are fused to obtain the final detection result.In this paper,training and testing are carried out on the commonly used UAV aerial image data set.The average detection accuracy of the adaptive cluster detection method in the Vis Drone2019 data set is 10.3% higher than that of the baseline detection method,and the detection accuracy of small objects is improved by 7.3% %.Using the real UAV images taken for testing,the results show that the adaptive clustering object detection method can effectively improve the detection accuracy of the model for the object in the clustered area,and the model has good generalization ability.