Research On Few-Shot Classification And Object Detection

Compared with deep neural networks,which often require a large number of annotated samples and repeated iterative training to learn the image features of new categories,humans only need one or several annotated examples to establish their cognition of new things and easily identify new categories.In order to make neural networks have the learning ability similar to human,a machine learning method called few-shot learning has become a research hotspot.This thesis mainly carries out research on image classification and object detection in few-shot learning.The main work and innovation points can be summarized as follows:(1)In order to make full use of the local features of images and improve the generalization ability of the model,a few-shot classification method based on local feature fusion is proposed.Firstly,the input image is processed into multi-scale grid blocks and then fed into feature extraction network to obtain local features.Secondly,a local feature fusion module based on Transformer architecture is designed to realize feature enhancement by integrating global information into local features and improve the model’s generalization ability.Finally,the Euclidean distance was used to calculate the distance between the feature vector of query set sample and the prototype of supporting set class.Compared with current advanced methods on the three data sets Mini Image Net,CUB and Tiered Image Net which commonly used in few-shot classification,the experiment results showed that,under 5-way 1-shot and5-way 5-shot settings,the average classification accuracy of the proposed method is2.04% and 1.68% higher respectively than the suboptimal method on the Mini Image Net dataset,and the average classification accuracy is 2.66% and 0.72%higher respectively than the suboptimal method on the CUB dataset.On the Tiered Image Net dataset,the proposed method is comparable to the state-of-the-art method in average classification accuracy.(2)In order to enhance the object detection network’s feature extraction ability,a few-shot object detection algorithm based on self-attention mechanism is proposed.Firstly,due to the difference of image features required by classification and detection tasks,it is proposed to separate the classification and detection branches of Faster RCNN prediction head for feature extraction.Secondly,the self-attention mechanism is introduced into the classification branch of the model to enhance the feature extraction and generalization ability of the model by fusing the local feature information.Finally,the proposed model was combined with TFA and De FRCN respectively,the detection results on the commonly used object detection dataset Pascal VOC and COCO were improved,and a large number of ablation experiments have also fully verified the effectiveness of the proposed method.