华中科技大学学报（自然科学版）2026，Vol.54Issue(5)：31-37,7.DOI:10.13245/j.hust.240839
基于困难样本挖掘的多尺度细粒度对象识别方法
Multi-scale fine-grained object recognition method based on hard sample mining
摘要
Abstract
To solve the issues of breed classification and individual identification in fine-grained cat and dog identification tasks,a lightweight visual model PetXNet was proposed.This model combined the YOLOv8 backbone network with a feature pyramid network(FPN)to achieve efficient feature extraction and multi-scale feature fusion.A triplet loss and hard sample mining strategy were adopted to improve the discriminative capability for different individuals.A phased training strategy was proposed to address different granularities of the dataset,and progressed from breed classification to individual identification through a coarse-to-fine and fine-to-coarse training approach.Based on public datasets,a self-constructed dataset was developed to fill the gap in individual verification data for pet biometrics.Experimental results show that PetXNet achieves high accuracy and generalization on both public and self-constructed datasets,and the model achieves 92.7％accuracy in breed classification tasks and 93.6％accuracy in individual verification tasks,verifying its effectiveness in fine-grained recognition tasks.关键词
困难样本挖掘/多尺度/特征提取/三元组损失/品种分类/个体验证Key words
hard sample mining/multi-scale/feature extraction/triplet loss/breed classification/individual verification分类
信息技术与安全科学引用本文复制引用
胡采瑄,马铭杰,李鉴,潘鹏,李国徽..基于困难样本挖掘的多尺度细粒度对象识别方法[J].华中科技大学学报（自然科学版）,2026,54(5):31-37,7.基金项目
装备预研教育部联合基金资助项目(8091B02072302) （8091B02072302）
国家自然科学基金资助项目(62272176). （62272176）
