基于深度学习的产品风格精细识别
Recognition method for fine-grained product styles based on deep learning
李雄 1苏建宁 2张志鹏 3祝铎 4鱼宝银5
作者信息
- 1. 兰州理工大学机电工程学院,甘肃 兰州 730050;兰州城市学院培黎机械工程学院,甘肃 兰州 730070
- 2. 兰州理工大学机电工程学院,甘肃 兰州 730050;兰州理工大学设计艺术学院,甘肃 兰州 730050
- 3. 兰州理工大学机电工程学院,甘肃 兰州 730050
- 4. 兰州城市学院培黎机械工程学院,甘肃 兰州 730070
- 5. 兰州理工大学设计艺术学院,甘肃 兰州 730050
- 折叠
摘要
为有效提取具有差异性的产品风格特征,提出一种基于复合学习通路的细粒度风格识别卷积神经网络(FSR-CNN).一是注意力学习通路,以残差结构为基础,采用串并结合的方式将坐标注意力、卷积块注意力和多头注意力嵌入其中,提出轻量化的混合注意力残差网络(HA-ResNet),用于抽取"专用特征".二是迁移学习通路,应用微调预先训练的GoogLeNet以扩充HA-ResNet模型容量,实现多感受野"通用特征"抽取.最后对二者输出的特征进行融合,并使用MLP分类器识别产品风格类型.在自行车头盔数据集上进行实验,并与其他经典深度卷积神经网络模型进行比较,实验结果表明FSR-CNN模型表现出较高的准确率和良好的稳健性,为产品风格精细检索与知识重用提供了一种新的模型算法架构.
Abstract
To effectively extract product style features with differences,a Fine-grained Styles Recognition Convolutional Neural Network(FSR-CNN)based on composite learning pipelines was proposed,which simulation integrated two key learning mechanisms of the human brain nervous system.From the attention learning pipeline,based on the residual struc-ture,the coordinate attention,convolutional block attention and multi-head attention were embedded in a string-parallel combination to form a lightweight Hybrid Attention Residual Network(HA-ResNet)for extracting"specialized features".From the transfer learning pipeline,the fine-tuning pre-trained GoogLeNet was used to expand the capacity of HA-ResNet model for extracting multi-receptive field"generic features".Finally,the output features of both were fused and the MLP classifier was used to identify the product style types.Experiments were performed on a self-built bicycle helmet dataset and compared with other classical deep convolutional neural network models.The experimental results showed that the FSR-CNN model exhibited higher accuracy and stronger robustness,which provided a new model algorithm architecture for prod-uct styles fine retrieval and reuse.
关键词
产品造型/风格识别/混合注意力/迁移学习/复合学习机制Key words
product form/style recognition/hybrid attention/transfer learning/composite learning mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据