基于孪生网络的自监督细粒度分类度量

Self-supervised fine-grained classification metrics based on twin networks

针对小样本间的细粒度分类中同种样本间不同个体的差异性不明显,导致特征难以提取的问题,设计了使用自监督的抠图式度量学习图像分类建模训练方法,在不增加数据集人工标注成本的基础上提高分类精度.首先,将数据集进行抠图处理,扩大检测目标在图像中的占比,突出样本特征.其次,通过比较网络结构模型及距离度量函数,择优对模型进行改进.最后,采用孪生网络的思想将样本集和验证集输入特征提取器中通过度量函数的计算得到两者相似度.选用的样本集为公开数据集CUB_200_2011和Standford Dogs Dataset,实验结果显示,提出的方法在性能和精度上得到了较好提升.

To address the problem of inconspicuous variability of different individuals among the same samples in fine-grained classification among small samples,which makes feature extraction difficult,a training method using self-supervised keying-based metric learning image classification modeling was designed to improve classification accuracy without increasing the cost of manual labeling of datasets.First,the dataset was keyed to expand the proportion of detection targets in the image and highlight the sample features.Secondly,the model was improved by comparing the network structure model and distance metric function,and the model was improved by merit.Finally,the twin network idea was used to input the sample set and the validation set into the feature extractor to obtain the similarity between them through the calculation of the metric function.The selected sample sets were the public dataset CUB_200_2011 and Standford Dogs Dataset,and the experimental results showed that the proposed method had a better performance and accuracy improvement.