Research On Improved Small Target Detection Based On Faster R-CNN Algorithm

Target detection is a very important research direction in computer vision.As one of the difficulties in target detection,small target detection exists in a large number of real scenes and is widely used in many fields such as automatic driving,intelligent medical treatment,defect detection and aerial image analysis.However,with the further development of deep learning,there are still some problems to be solved in object detection,which affect the development to a certain extent.One problem is that the detection of small objects is not effective.In order to effectively solve the problem of poor detection performance of small targets,this paper discusses the reasons for low detection accuracy of small targets based on Faster R-CNN algorithm,and puts forward three effective improvement methods to improve detection efficiency and accuracy.The three methods are data enhancement,feature fusion network improvement and post-processing algorithm improvement to improve the small target detection algorithm,and the effect of the improvement is verified.(1)A data augmentation method was proposed to solve the problems faced by small targets,such as low resolution,limited extractable features,limited sample size,and uneven distribution.In this paper,a variety of single sample data enhancement methods and multi sample data enhancement methods are used to increase the number of small object samples,not only make the characteristics of small objects more obvious,but also improve the anti overfitting ability of the network to a certain extent,and four data enhancement methods are selected from them,which make the detection accuracy of small objects improve most.Finally,the experiment proved that using data augmentation can greatly improve the detection accuracy of small objects.(2)An improved feature fusion network was proposed to solve the problem of excessive convolution of shallow features caused by feature fusion networks.This article replaces the FPN structure with the PANet structure and adds a new feature fusion structure called CBAM attention module to the structure.This not only optimizes the path of shallow features during feature fusion,allowing for the preservation of shallow features that are conducive to small target detection,but also adds a CBAM attention mechanism that can selectively suppress the background.Finally,the effectiveness of the model in improving the accuracy of small object detection was verified through experiments.(3)An improved post-processing module has been proposed to address the considerable limitation of traditional NMS algorithms being unable to obtain global optimal solutions.This article designs a self attention post-processing algorithm by learning to remove redundant detection boxes,which can output a unique detection box for each target in the image.The self attention post-processing algorithm first divides the detection frame into different clusters belonging to different targets using DBSCAN clustering.Secondly,the coordinates,area,category scores,and feature vectors of the fusion detection box are used to represent the features of the suggestion box.Then,a self attention mechanism is used to fuse the feature information of the detection box for the same target.The fused feature information is input into the GRU network for training,and a unique detection box is output for the target in the image during detection.Finally,the effectiveness of the model was verified through experiments.