基于参数优化残差网络的皮革缺陷分类

针对皮革缺陷分类存在误判、成本较高及目前关于皮革缺陷的研究主要是针对皮革做缺陷检测,未进行缺陷分类的问题,采用一种参数优化的残差网络来实现皮革缺陷的自动分类。首先通过多层卷积、池化操作进行特征提取,并引入残差模块解决深层网络的梯度消失问题;然后依据所提取特征进行缺陷分类;最后根据皮革数据集优化关键网络参数,使用数据增强方法对数据集进行扩充,有效避免了网络模型因样本不足易产生过拟合的问题。实验结果表明该方法可对皮革缺陷进行有效分类,分类精度达到92.34%。

Research on Classification of Leather Defects Based on a Parameter-optimized Residual Network

The current research on leather defects is mainly aimed at defect detection without classification of defect types. Also, manual classification has disadvantages including misjudgment and high cost. To address these issues, this paper presents a parameter-optimized residual network to realize automatic classification of the leather defects. Firstly, features are extracted by multi-layer convolution, pooling operations. And a residual module is introduced to address the issue of gradient disappearance in the deep network. Then the defect classification is carried out according to the extracted features. Finally, the key network parameters are optimized according to the leather data set. In addition, a data enhancing method is utilized to augment the dataset, which effectively avoids the over-fitting problem caused by insufficient data volume. The experimental results indicate the proposed method can accomplish the classification according to the defect types and the classification accuracy can reach 92.34%.