摘要
针对某一类别商品图像的精细分类,研究并实现了深度学习中的卷积神经网络方法。所设计的卷积神经网络由2个卷积层、2个亚采样层及1个完全连接层组成,特征平面的神经元只对其感受野的重叠区域做出反应,由反向传播算法调整网络参数最终完成学习任务。通过鞋类图像的精细分类实验表明,该方法平均分类正确率可达91.5%。
For the fine-grained classification of product image, convolutional neural network was explored and imple- mented as one deep learning method. The designed convolutional neural network consisted of two convolution layers, two subsampling layers and one fully connected layer, where the individual neurons were tiled to respond to overlapping regions in the visual field. The learning task was accomplished through adjusting the network parameters with back propagation algorithm. The proposed method has achieved an average accuracy of 91.5% in the fine-grained classification tests for the shoe images.
出处
《山东科技大学学报(自然科学版)》
CAS
2014年第6期91-96,共6页
Journal of Shandong University of Science and Technology(Natural Science)
基金
辽宁省教育厅高等学校科学研究项目(L2014174)
关键词
卷积神经网络
商品图像
精细分类
亚采样层
convolutional neural network
product image
fine-grained classification
subsampling layer