结合自监督对比与元迁移的小样本图像分类算法  

Small Sample Image Classification Algorithm Combining Self-supervised Contrast and Meta-migration

作　　者：贾冰 王军号[1] JIA Bing;WANG Junhao(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《兰州工业学院学报》2024年第5期8-14,共7页Journal of Lanzhou Institute of Technology

基　　金：国家自然科学基金(61300001)。

摘　　要：针对元训练过程易忽略图像中包含的自监督信息而导致特征提取不充分,以及模型泛化能力差的问题,提出结合自监督对比与元迁移的小样本图像分类模型。预训练阶段提出自监督全局和局部联合对比损失对特征编码器进行训练,使其学习到更丰富的可迁移的先验知识;元训练阶段进一步对样本在特征空间中易产生偏离的问题,提出原型中心度量网络来优化特征空间,使同类样本特征分布更紧凑;利用基于余弦相似度度量的余弦分类器计算各类别中心与查询集图像的相似度;在Mini-ImageNet和Tiered-ImageNet数据集上与当前主流模型进行分类效果对比实验。结果表明:与baseline相比,所提出的模型准确率在Mini-ImageNet数据集5-way 1-shot和5-way 5-shot任务中分别提升了3.14%和4.09%,在Tiered-ImageNet数据集的两个任务中分别提升了2.98%和3.73%。To solve the problem that self-supervised information contained in images is easily ignored during meta-training,which leads to inadequate feature extraction and poor model generalization ability,a small sample image classification model combining self-supervised comparison and meta-migration is proposed.Firstly,in the pre-training stage,a self-supervised global and local joint comparison loss is proposed to train the feature encoder,so that it can learn richer transferable prior knowledge.Secondly,in the meta-training stage,to solve the problem that samples are prone to deviation in the feature space,a prototype center metric network is proposed to optimize the feature space and make the feature distribution of similar samples more compact.Finally,the cosine classifier based on the cosine similarity metric is used to calculate the similarity between the center of each category and the query set image.The classification effect is compared with the current mainstream model on the Mini-ImageNet and Tiered-ImageNet datasets.The results show that:Compared with the baseline,the accuracy of the proposed model is improved by 3.14%and 4.09%in the 5-way 1-shot and 5-way 5-shot tasks of the Mini-ImageNet dataset,respectively,over the two tasks of the Tiered-ImageNet dataset,is improved by 2.98%and 3.73%,respectively.

关 键 词：图像分类 小样本学习 自监督学习 对比学习 度量网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]