Study On Image Classification In Few Shot Learning Based On Deep Neural Network

In the last ten years,with the development of computating devices and the massive data collected,deep learning and machine learning related algorithms have reached remarkable success in image processing.However,in most real-world application scenarios,it is an impossible task to collect a large amount of training data,such as the data collection of a rare case in a medical scenario.In addition,marking a large amount of data also requires a large amount of time and manpower.Therefore,image classification algorithm under a small number of samples has gradually become a popular research direction,also known as few shot learning.Few shot learning aims to train a model with stronger generalization ability through a small number of samples.This thesis summarizes and analyzes the existing few shot learning algorithm.In order to improve the generalization ability of few shot learning model,the overfitting problem and the multi-scale representations ability in few shot image classification,this thesis proposes a new multi-branch network structure.The details are follows:1.A network framework based on multiple branches is proposed.This framework takes advantage of the multi-scale characteristics of deep convolutional neural networks,which add the feature maps of adjacent scales to generate a basic branch unit.Then,for each branch unit,the label of the corresponding training sample is used to make a separate constraint.The multi-branch structure is a new kind of deep supervision structure,which can directly optimize the shallow layer of the network,thus enhancing the anti-overfitting ability of the network.2.The collaborative learning loss function and soft target loss function are proposed.Based on the multi-branch framework in Part 1 above,this thesis further proposes collaborative learning loss function and soft target loss function.Collaborative learning loss function works among the branch units,using the improved KL divergence distance(Kullback Leibler divergence)to minimize the distance of prediction probability between the shallow branch unit and deep branch unit.And the shallow branch unit can mimic the optimization manner of deep branch unit,which is beneficial to the representation ability of the network.In addition,in order to enhance the feature representation ability of the head layers,the soft target loss function makes use of the prediction information of the deep layers,so that the head layers can get more target information,so as to better optimize the network.The collaborative loss function and soft target loss function can enhance the generalization ability of the network in scenarios with fewer samples,so that the network can learn more discriminative features with fewer samples.