Research On Optimization Algorithm For Small Target Detection Based On Single-stage Detecto

With the continuous development of intelligent technology,object detection algorithm based on deep learning is gradually widely used in many fields.Compared with traditional algorithms,the detection algorithm based on deep learning can deal with higher requirements and more difficult image detection tasks.In recent years,there are more and more complex and rich application scenarios,and the requirements for the real-time detection performance of the algorithm are gradually improved.The one-stage detection algorithm in the form of end-to-end detection has become a hot spot in the detection field because it can have very high detection accuracy with less parameter calculation.However,the one-stage detector also has some disadvantages,its detection effect on small objects is relatively poor,and there are often false detection and missed detection of small objects.Therefore,how to improve the detection effect of one-stage detection algorithm on small objects has very important research significance.In order to solve the problem of imbalance between large and small object features caused by relatively few small object features in the data set,a data enhancement method based on small object feature enhancement is proposed in this paper.Firstly,through the small object interception method proposed in this paper,the small object image is extracted from the existing object detection data set and the small object data set is established.Then,through the improved Cut Mix method,the images in the small object data set are mixed into the data set to be enhanced to achieve small object feature enhancement.The characteristic of data enhancement method based on small object feature enhancement is that it is a data enhancement method that can cross the image dimension,distinguish the object size in the image,and enhance only the object with small scale.SSD algorithm adopts the detection method of shallow feature learning small object and deep feature learning large object.However,because the shallow network is used for feature summary,there are some problems,such as the amount of calculation of shallow convolution is not as large as that of deep features,the convergence speed of attention area of shallow feature unit is relatively slow,and the parameter update is more affected by the disappearance of gradient.In view of the above problems,this paper proposes two model frameworks: SFE-SSD and ASL-SSD.SFE-SSD framework enhances the shallow feature map through a new multiscale feature fusion method to solve the problem of insufficient convolution computation in the summary of shallow network features.ASL-SSD framework improves the ability to summarize the characteristics of small objects by accelerating the updating speed of shallow network parameters.Firstly,the feature weight of the feature unit with objects in the attention area is increased through the ALAM spatial attention module,and the convergence speed of the attention area with high weight is improved.Then the residual multiplexing module RM is added to the shallow network to improve the ability of the shallow network to suppress the disappearance of gradient and speed up the update of shallow parameters.The experimental analysis on NUWPU VHR-10,PASCAL VOC2007 and MS COCO2014 data sets shows that the enhancement algorithm in this chapter and the model in this paper are effective for the one-stage object detector to improve the detection effect of small objects.