Adaptive deep few-shot learning

Abstract

Machine learning (ML) techniques have been successfully implemented in many fields with the absence of sufficient and high-quality data, such as computer vision (CV) tasks. However, the performance of ML techniques may be derogated due to the presence of insufficient data. Regarding the problem of this aspect, Few-shot learning (FSL) has been developed recently to solve the problem caused by the mismatch between the quantity of current dataset and ideal dataset respectively. In this dissertation, we study the architecture of various convolutional neural networks like Visual Geometry Group (VGG) and residual neural networks (ResNet) and implement FSL in image classification based on different VGG neural networks and ResNets. The classification accuracy ranges from 22.32% to 98.72% with different neural networks as the feature extraction network. The result of most experiments is above 92.01%, which shows the efficient image classification ability of ResNets and VGG neural networks in FSL.