The Research On Few-Shot Object Detection Method

With the development of information technology,deep learning has been widely used.Object detection methods based on deep learning have achieved advanced research results in many fields.Deep learning requires a large amount of labeled data to support the performance of the model.Therefore,how to use a very small amount of data to achieve high-precision object detection has become a new research hotspot this year.In the case of few data,the deep neural network is prone to problems such as difficult convergence,over fitting and poor generalization performance;And because the sample data is too small,the information is lost in the process of feature extraction,and the insufficient weighting of feature fusion will have a great impact on the performance of model detection;And in the case of feature redundancy,the model will be too complex to fit the model parameters,resulting in low detection accuracy.In view of the above problems,this paper carries out research on few shot object detection,and the main work is as follows:In view of the feature loss in the process of feature extraction from high resolution to low resolution in few shot object detection,and the insufficient fusion of support set features to query set features,this paper proposes a few shot object detection algorithm based on meta learning network model.Based on the idea of meta learning,the algorithm uses single-stage object detection as the basic detection model,redesigns the network model,and realizes object detection in small sample scenes.Firstly,the algorithm adopts a high-resolution feature extraction network to ensure that the extracted features maintain high resolution;Secondly,in order to reduce the interference of irrelevant information such as background,the model constructs parallel channels to obtain the information of key areas of the image,and introduces spatial attention model to increase the weight of key information;Finally,a transformer like method is used for feature addition and fusion.It uses self-attention to enhance the features of meta knowledge,and then uses cross attention for feature fusion.Experiments show that the algorithm has achieved good detection results on the public data set of small sample target detection,especially when the number of pictures exceeds 3,the detection accuracy is significantly improved.In few shot object detection,due to too little sample data and too many model parameters,it is easy to have unfitting and generalization performance when directly using a small amount of data to fine tune the network model.Based on the fine-tuning network model,this paper designs and implements a few shot object detection model that can reduce feature redundancy and enhance generalization.When extracting features from the backbone network,the redundant information in the model will increase the difficulty of fine-tuning the model.This paper uses the feature splitting and fusion structure in the residual split(RS)structure to reduce the redundant features in feature extraction.In the network training of few shot object detection,the classifier using cross entropy loss function is prone to over fitting,because there are a large number of parameters in the classifier that need to be optimized.In this paper,Pearson distance function is used to calculate the classification loss of the model,that is,non-parametric method is used to reduce the number of parameters,so as to prevent over fitting.Experiments show that the few shot object detection model designed in this paper is better than other models on Pascal VOC and FSOD datasets.