Image Classification And Detection With Few-shot

With the development of computer vision technology,image classification and detection technology based on deep learning has been widely applied in many application fields.At present,the deep learning method relies on a large number of data for training,while data collection,labeling and sorting require huge human,material and finanical resources.However,the problem of insufficient data is still imperfectly solved in some areas.Accordingly,few-shot learning provides a feasible way to solve the above problems.But few-shot classification methods are complex and lack of practicability.In addition,few-shot image detection methods are just start on,and there is no general scheme at this stage.In this paper,we simplify the model of few-shot classification to increase its practicability,and then realize the general framework of few-shot detection.The main contents are as follows:1.Aiming at solving the problem of the poor practicability of the current model in few-shot classification,the method of few-shot image classification based on metric learning and feature aggregation(Simplified Metric Learning Framework,SML)is designed and implemented,and the evaluation standard based on model memory ability(Mean Memory,MM)is proposed.The training feature extraction network of batch training mode,combined with feature aggregation,the end-to-end training is realized.It gets rid of the disadvantages of new tasks,which requires fine tuning or retraining,and improves the practicability of the model.The experimental results show that SML has higher accuracy and better transfer ablility,and MM is more practical and comprehensive for the evaluation of few-shot classification models;2.A group of general modules for few-shot image detection are proposed,which are global context attention mechanism module(Global Context Similarity Module,GCSM)and aggregation module.GCSMDet consists of this group of modules and the general object detector,which can effectively extract attention information including supporting images and querying images,and learn the spatial similarity between support images and query images.The experimental results show that the attention mechanism module is universal.Compared with the recent methods,GCSMDet not only contains less parameters,but also has higher accuracy;3.We build a demonstration system for few-shot classification and detection.The system can transmit the image to the server,receive and display the visual results of few-shot classification and detection.