基于流形混合增强的小样本图像分类算法

A Few-shot Image Classification Algorithm Based on Manifold Mixing Augment

针对训练样本不足,导致训练模型容易过拟合的问题,提出了一种流形混合增强方法。首先,将经过简单数据增强处理的样本输入网络,在训练过程中,随机选择一个输出特征层;然后,从训练集中随机抽取训练样本,按照混合参数进行不同方式的混合增强处理后,再送入下一个特征层,以此来缓解模型过拟合的问题。最后,针对传统神经网络获取长距离依赖关系困难的问题,融入一种光谱非局部块,使得特征提取网络可以更稳健、灵活地捕获远程依赖关系,以此来优化网络,提高网络性能。实验结果表明,在CIFAR-FS、CUB200和miniImageNet三个数据集及跨域分类上进行5-way 1-shot任务的平均与基准算法准确率对比,所提算法准确率有较高提升,验证了所提算法的有效性。

A manifold mixing augment method to deal with the problem that the training samples were insufficient,which could make the training model being easily overfitted,was proposed.Firstly,the dataaugmented samples were input into the network.During training,an output feature layer was randomly selected.Then,the training samples were randomly selected from the training set for mixed enhancement processing in different ways according to the mixed parameters,and then sent to the next feature layer to alleviate the problem of overfitting.Finally,in view of the difficulty of obtaining long-distance dependencies with traditional neural networks,a spectral non-local block was incorporated,which enabled the feature extraction network to capture long-distance dependencies more robustly and flexibly,so that the network was optimized and the network performance was improved.The experimental results showed that the classification accuracy of the 5-way 1-shot task on the CIFAR-FS,CUB200 and miniImageNet datasets and cross-domain classification were all improved compared with PT+MAP algorithm,verifying the effectiveness of the proposed method.