Research On Small Target Detection Technology Based On Image Enhancement Method

At present,target detection algorithms represented by deep learning,such as: You only look once(YOLO),Single Shot Multi Box Detector(SSD)and other single-stage target detection algorithms have demonstrated high detection rates and stable detection effects However,this type of detection algorithm is used in some specific small target detection applications,such as crowded face targets in the station square,and bird targets in the distance of the terminal.Because the relative and absolute sizes of these targets are very small,the traditional The detection rate of the detection algorithm for these small targets is only half of that of ordinary targets.Because of the low detection rate and high false detection rate,traditional detection methods cannot meet the requirements of these application scenarios.In response to this problem,this paper proposes a small target detection technology based on image enhancement methods to solve the problem of low accuracy of small target detection in such airport scenes.The main content and contributions of its research are as follows:(1)In the small target sample,because the target to image ratio is small,it cannot be directly used for detection,so it is necessary to transform the sample data.Image segmentation is used to increase the proportion of small targets and retain the information of the small targets,which is used to solve the problem of the small proportion of the target to the original image.(2)In the collected small target sample data,most of the targets are small and fuzzy,and cannot be directly used for the training of target detection algorithms,and need to be enhanced.A method for enhancing the resolution of small targets based on Super-Resolution Using a Generative Adversarial Network(SRGAN)is proposed to improve the resolution of small targets and enhance their features and context information.(3)The use of image block and SRGAN is to enhance the characteristics of small targets in the sample,making it easier for the algorithm to detect small targets,and the enhancement effect depends on the appropriate block ratio and SRAGN magnification parameters.A method for evaluating the optimal parameters of image enhancement is proposed,which mainly uses image quality evaluation and normalized target detection efficiency to evaluate the effect of image enhancement and obtain the optimal parameters of image enhancement.The method proposed in this paper has carried out experiments on the detection of flying birds in airports.The average accuracy and average accuracy(AP)of small target detection are 90.26% and 67.75%,which are compared with the traditional YOLO v4 and the small target detection of airport flying birds that directly introduce Context information.In comparison,the accuracy rate and average accuracy are increased by 15.70% and 16.12%,respectively.