基于深度学习的长尾数据集分类精度提高的研究

Research on Improving Classification Accuracy of Long Tail Dataset Based on Deep Learning

基于深度学习的轮胎缺陷检测过程中,由于轮胎数据集为长尾数据集,某些缺陷类别的轮胎图片数量较少,造成此类缺陷的数据分类精度不高。本文在双边分支神经网络(BBN)分类算法的基础上,通过改进特征提取网络的结构、损失函数和权衡参数提高轮胎数据集中尾类数据分类精度。以主流数据集长尾CIFAR-10、长尾CIFAR-100和本实验室自制轮胎数据集为测试数据进行实验分析,实验结果表明数据集的尾类小样本数据的分类精度提高约10%。

In the process of detection of tire defects based on deep learning,the tire data set is a long tail data set,so the number of tail defect images in the data set is small,resulting in low classification accuracy of such defects.This article is aimed at a classification algorithm based on BBN(Bilateral-Branch Network)bilateral branch neural network.By changing the structure of the feature extraction network,the loss function and the optimization of the training speed,the optimized network can classify the tail data in the tire data set with higher accuracy.Through training on the mainstream data sets Long-tail CIFAR-10,Long-tail CIFAR-100 and our laboratory′s self-made tire data set,the experimental results show that the classification accuracy of the small tail sample data of the data set has been improved10%.