Research On Few-Shot Detection Algorithm Based On Two-Stage Model

In recent years,in the field of few-shot image recognition,more research has been devoted to issues,but in fact,detection issues are more practical and valuable.Currently,few-shot target detection algorithms are in the early stages of research,and the research on few-shot detection algorithms is of great significance.This thesis improves the few-shot detection framework based on the two-stage detection model.Firstly,the detection head of the second stage of the two-stage detection model is improved,and an improved detector based on the DML(Distance Metric Learning)subnet is designed.The category is evaluated by the distance between the ROI(Region of Interest)output vector and a representative class in the embedded space with multiple mixed models,and then a comparison network is introduced to add a detector as a parallel branch of the DML subnet,which is used to improve the similarity between the same kind and the difference between different types.The ROI pooling is improved to ROI Align to solve the problem of misalignment between the original image and the feature map in the traditional two-stage detection model.Secondly,improve the feature extraction network of the first stage of the two-stage detection model,and design an improved feature extraction network based on data augmentation.A series of high-quality virtual labeled data is generated through three steps: candidate frame generation,label correction,and border correction.A change network is designed to accept all valid data that can be changed and generate new and changed training data for training,thereby increasing the quantity and quality of new types of data.In addition,in order to better extract features from images,ResNET-FPN,a multi-scale feature pyramid,is used to replace the original backbone network,Integrate features at all levels to enhance semantic and spatial information.The few-shot learning method based on which the two different stages of improvement modules are based is also considered from different dimensions.The improved detector is considered from the perspective of distance metric learning,while improving the feature extraction network is the most fundamental problem of few-shot learning-the lack of data sets to consider.The two improved modules can be used separately with traditional two-stage detection models,or can be combined into a unified framework.Experiments have shown that the performance of each module and the combined framework is superior to traditional algorithms.Compared to older algorithms,mAP has an average improvement of 69%,and compared to advanced algorithms,it has an average improvement of 4%.At the end of this article,a few-shot detection system is designed,in which multiple unmanned aerial vehicles cooperate to perform few-shot detection tasks.The few-shot detection system is divided into two modules: software and hardware.The improved few-shot algorithm in this chapter is deployed to the hardware platform,and the software is deployed to the ground PC server.The few-shot detection system is completed through mutual communication between the two modules.