Research On Few-shot Object Detection Algorithm

Few-shot object detection refers to the classification and positioning of the target in the image by training a small number of labeled samples.At present,the main object detection algorithms are based on a large number of data annotation.And a large number of data collection and data annotation are often very difficult in many practical scenarios.So few-shot object detection technology has very important application value and has attracted more and more researchers'attention.This paper will combine the method of few-shot learning and deep learning to carry on the research to the few-shot object detection technique.In this paper,we propose a new solution to some problems existing in the current few-shot object detection algorithm.In order to solve the problem that the target category is difficult to distinguish from the other category of non-target and the effect of subsequent subdivision is poor,a few-shot object detection model based on attention mechanism and feature representation metric learning is designed.The model uses the attention mechanism to introduce the information of the target category into the RPN and designs the Embedding module and the category decision method based on metric learning.In order to solve the problems of poor generalization performance,easy overfitting and insufficient richness of data scale,a few-shot object detection network based on transfer learning and weight initialization is designed.The model uses a multi-scale fusion strategy and data enhancement strategy to design an initialization weight training method based on task generality.And a knowledge transfer regularization and background suppression regularization method based on transfer learning are also designed.This paper carries on the experimental verification in the data sets MS COCO and Image Net.On the Image Net test set,this method is 1.3%higher than the SOTA method on the AP₅₀index.The model trained on the FSOD dataset increases the AP₅₀/AP₇₅index by 0.2%/6.4%compared with the model fine-tuned on the MS COCO.The fine-tuned model based on FSOD data increased the AP₅₀/AP₇₅index by 2.6/2.7%.On the MS COCO test set,the trained model is a big step ahead of the classical Meta R-CNN and is 5.2%/8.1%/5.7%ahead of the AP/AP₅₀/AP₇₅index,respectively.The results of training in FSOD data sets have been further improved and the AP/AP₅₀/AP₇₅indicators have been increased by 7.4/11.7/8.7%respectively.Through the research of few-shot object detection algorithm,it can greatly reduce the workload of data collection and high quality data annotation in real scenes,and greatly promote the landing and popularization of object detection algorithm in each scene,which has important practical significance.