Research And Application On Classification And Detection Methods For Few Shot Learning

In recent years,with the development of deep learning,image classification and object detection based on neural network have made breakthrough progress.These deep neural network-based classifiers and detectors often need a large number of training samples to achieve ideal detection results.However,in many cases,such as medical field,aerospace field,endangered wildlife,etc.,the number of samples is far smaller than the requirements of these classifiers and target detectors.When dealing with small sample data,neural networks with a large number of training parameters often have the problem of overfitting.This paper focuses on the small sample data classification and detection algorithm based on deep learning,analyzes the advantages and disadvantages of different methods for small sample data,analyzes the two aspects of image classification and target detection,and applies it to Parkinson’s handwritten data set classification,Mars terrain detection and other fields.The main work of this paper is as follows:(1)Since the training data set is small,transfer learning is usually used to solve the problem of small training data set.In this paper,the migration based on parameter learning to improve the network generalization ability,avoid network fitting,at the same time puts forward a new method to improve the network training speed,on the basis of the training model,fixed convolution layer parameters,by adding the new parameter in each layer to reduce the number of learning parameters,which can achieve the purpose of quickly improve the network training speed.In addition,Earth Mover’s Distance(EMD)is used as a metric to calculate the structural Distance between dense image representations to determine the correlation of images.EMD can be inserted into the network as a layer for end-to-end training.Experimental results show that the proposed algorithm improves the classification accuracy of Parkinson’s handwritten images.(2)Aiming at the lack of central point information of small sample data,this paper proposes a new small sample target detection method based on the combination of high and low resolution features.The high resolution representation module is used to generate heat maps containing rich high-level characteristic information.Meanwhile,in order to enrich the information of the target center,the target region is input into the low resolution representation module and combined with the high resolution heat map,the location and size of the target are regressed.Experiments on Pascal VOC2007/2012 data set show that the high and low resolution representation module can significantly improve the detection accuracy.(3)In order to improve the stability of small sample detector,different from existing small sample target detection methods,such as improving detection classification head and cutting target image as auxiliary data,this paper emphasizes that focusing on target center point can solve the generalization problem of small sample detector recognizing new categories.A new attentional mechanism based on target center auxiliary feature circle graph is proposed to improve the performance of small sample target detection.By selecting an auxiliary feature circle graph with the target center as the center and the minimum length and width as the radius,it was added to the anchor-free Centernet network architecture as a soft attention mechanism.Several experiments on Pascal VOC2007/2012 data set show that the proposed soft attention mechanism can achieve the most advanced level in small sample target detection,and has been successfully applied to Mars terrain detection mission,demonstrating the effectiveness of the proposed method.