摘要
针对现有公开体育器材数据集较少且种类有限的缺点,构建了一个新的数据集SED(Sports Equipment Dataset),该数据集具有分布均衡、多样性高、背景丰富等优点.对于多类别小规模数据集,单一模型预测效果不能达到预期的准确率,因此在构建SED数据集基础上,提出了一种模型融合与迁移学习相结合的方法.选取ResNet50和InceptionV3作为特征提取器,将2个模型提取的特征融合输入到全连接层再实现分类.同时利用迁移学习的方法优化模型参数,进一步提高模型精度.实验结果表明,在涉及69类体育器材图片分类任务中,准确率达到85%,对体育器材图片分类具有较好的效果.
Aiming at the drawback that the current publicly-available sports equipment dataset are insufficient and limited in types,a new dataset SED(Sports Equipment Dataset)is constructed,which has the advantages of balanced distribution,high diversity,and rich background.For multi-category small-scale dataset,the prediction effect of a single model cannot reach the expected accuracy.Therefore,based on the construction of the SED dataset,a method combining model fusion and transfer learning is proposed.ResNet50 and InceptionV3 are selected as feature extractors,and the features extracted by the two models are fused and input to the fully connected layer to achieve classification.Simultaneously,transfer learning is used to optimize model parameters to further improve model accuracy.Experiments show that in the classification tasks involving 69 types of sports equipment pictures,the overall accuracy of the algorithm reaches 85%,which has a good effect on sports equipment image classification.
作者
石瑞
艾山·吾买尔
早克热·卡德尔
王中玉
杰恩斯艾力·努尔达艾勒
SHI Rui;AISHAN·Wumaier;ZAOKERE·Kadeer;WANG Zhong-yu;JIEENSIAILI·Nuerdaaile(College of Information Science and Engineering,Xinjiang University,Urumqi 830046,China;Xinjiang Laboratory of Multi-Language Information Technology,Xinjiang University,Urumqi 830046,China)
出处
《东北师大学报（自然科学版）》
CAS
北大核心
2022年第4期54-63,共10页
Journal of Northeast Normal University(Natural Science Edition)
基金
新疆维吾尔自治区自然科学基金联合基金面上项目(2021D01C081)
国家重点研发项目子课题(2018YFB1403202)
天山创新团队计划(2020D14044).
