Rotating Small Target Detection Based On Improved Faster R-CNN

Nowadays,target detection technology based on deep learning has made great progress,and it has been widely used in fileds such as unmanned driving,military technology,intelligent disease diagnosis and security systems.However,the target detection algorithms based on Faster R-CNN have poor performance in feature extraction of small-scale target,and the performance of target detection based on rotating frame is low.According to the characteristics of smallscale targets and rotating targets,this paper proposes an improved the detection technology of small rotating targets based on Faster R-CNN,and achievesd good detection results.The research work of this project mainly includes two aspects:Firstly,a small-scale target detection algorithm based on improved feature coding and adaptive pooling layer is proposed.In the first step deformable convolution and bottom-up connection computation are used to enhance the bottom position information in of feature extraction structure.While enhancing the sampling of the receptive field,it further increases the detection effect of small-scale targets.In the second step,Gaussian function is introduced to calculate the region overlap frame threshold in linear non maximum suppression,which improves the detection accuracy of dense regions.On this basis,the adaptive pooling layer is used to sample the region of interest,which effectively avoids the candidate frame offset caused by quantization calculation of pooling layer.Secondly,this paper proposes a method of rotating target detection based on region overlap and rotating anchor frame.Different from the horizontal border labeling methods with Faster R-CNN,this paper proposes a regression method based on rotation anchor frame and rotation region overlap,which reduces the numerical oscillation of rotation angle regression and improves the detection accuracy of rotating target in dense areas.Then,the paper analyzed,the inadaptability of smooth L1 loss function to the calculation of rotation angle and proposed a method introduced the regional overlap degree in the loss function,the super parameter and regional overlap degree are used to improve the accuracy of rotation angle regression.A series of experimental results show that the proposed algorithm can enhance the detection effect of rotating targets and small-scale targets.Aiming at the effectiveness analysis of the target detection algorithm proposed in this paper,small-scale target data set(Tiny Person)is used for comparative experiments.The experimental results show that the small-scale target detection algorithm based on improved feature coding and adaptive pooling layer is effective,and the average detection accuracy reaches 57.7%.The experimental results show that the average detection accuracy of the proposed algorithm has reahced 80.3%,which effectively improves the detection accuracy of rotating small-scale targets.