基于卷积神经网络与显著性特征的皮革缺陷检测

Inspection Method of Leather Defect Based on Convolutional Neural Network and Salient Feature

针对目前皮革缺陷形态多样、检测难度高的问题,提出卷积神经网络与显著性特征组合的检测方法。以9种常见的皮革缺陷为检测目标,应用卷积神经网络方法对其进行初步检测,结果表明此方法对其中几种缺陷的检测准确性不够。对皮革缺陷的几何和灰度特征进行数理统计,归纳出缺陷显著性特征描述,提出以卷积神经网络为主、显著性特征为辅的组合检测方法。通过试验验证得出,组合检测方法的准确率可达90%以上,相比卷积神经网络法,检测准确率有所提升且平均处理时间的增幅很小,可满足实际皮革缺陷检测需求。

In order to solve the problem of leather defects with various forms and high detection difficulty,a method combining convolutional neural network and salient feature was proposed.Taking nine common leather defects as detection targets,the convolution neural network method was applied.It was found that the accuracy of detection was not qualified for several kinds of defects.Thus,the mathematical statistics of geometric and gray features of leather defects were calculated and the salient features were summarized.Then a combined detection method with convolutional neural network as the main and salient features as a complement was proposed.The experimental results show that the classification accuracy of the method is more than 90%.Compared with the convolutional neural network method,the detection accuracy was improved while the increase in average processing time was small,which could meet the actual needs of leather detection.