Research On Deep Learning Based Rotating Object Detection Algorithm

With the rapid development of deep learning and computer vision technology,nowadays the integration of aerial image processing technology and object detection technology is becoming more and more mature.It is widely used in various fields,such as military field: object reconnaissance,enemy identification,base survey,etc.,and civilian field: land resource survey,construction planning,traffic flow supervision,etc.Also object detection is facing many challenges and difficulties.Since the orientation of targets in aerial images is always arbitrary,they are called rotating objects.Due to the special characteristics of rotating objects,such as small objects,different shapes and sizes,dense and irregular arrangement among objects,etc.,there are also certain limitations of the object detection algorithm based on the horizontal frame,and there are still problems such as features not aligned between the object and the detection frame.For the above problems,the following innovations and improvements are made in this paper:First,for the arbitrary direction and dense arrangement of rotating objects in aerial images,this paper improves the horizontal detection frame in the object detection algorithm model to a rotating detection frame,which can better fit the shape of the rotating object itself,reduce the proportion of background in the detection frame,improve the visual perception,and enhance the detection effect.Secondly,for the problems of small rotating objects and complex backgrounds in aerial images,this paper improves the backbone network in the object detection algorithm model by selecting a large scale feature layer with a feature pyramid structure for channel fusion as the detection head output of small objects,the large scale feature layer contains more shallow semantic information and retains more detailed features of the object,which can overcome the complex backgrounds on small object detection It can overcome the interference of complex background on small object detection and improve the detection effect on small objects.Then,for the misalignment problem between rotating objects and anchor frames in aerial images,this paper adds a feature alignment structure and a direction detection structure to the detection head.The feature alignment structure extracts the alignment features between objects and anchor frames to form an alignment feature layer;the direction detection structure uses a rotating convolution kernel to extract the input alignment feature layer for the task of classification and regression,which effectively alleviates the feature unalignment problem of objects and improves the detection accuracy.Finally,the YOLOv4 algorithm model is selected for real-time requirements,and the performance of YOLOv4 is further improved by using a high-performance backbone network,an improved neck structure,an improved detection head,and an improved loss function.In this paper,DOTAv1.0 aerial image dataset is selected for multiple ablation experiments to verify the authenticity and effectiveness of the improved method in this paper.