基于多任务特征层共享的表面缺陷检测

Surface Defect Detection Based on Multi-task Feature Layer Sharing

在基于卷积神经网络的工业表面缺陷检测中,往往将缺陷检测任务视为单任务进行学习。针对单任务学习数据来源单一、样本量缺乏的问题,提出了一种基于多任务特征层共享的表面缺陷检测方法(Feature-shared Multi-task Network,FSMTNet)。该方法首先对原图像数据分别进行高斯滤波和canny算子边缘提取,处理后的两组图像数据和原数据视为三个相关联的任务;三组图像分别输入以AlexNet为主干的网络,通过一种共享单元(Shared unit)进行任务间的特征共享;并在一系列可学习权重的作用下,自主学习出最佳共享组合。在不进行任何数据增强的前提下,该方法达到了96.111%的平均准确率,对比传统方法提升6%,子类准确率最大提升27%。实验结果表明,基于多任务特征层共享的缺陷检测网络性能优于同类单任务网络。

The defect detection task is often regarded as a single task for learning in industrial surface defect detection based on convolutional neural networks.Aiming at the problem of single source of single task learning data and lack of sample,this paper proposes a surface defect detection method based on multi-task feature layer sharing(Feature-shared Multi-task Network,FSMTNet).The method first performs Gaussian filtering and canny operator edge extraction on the original image data,and the processed two sets of image data and original data are regarded as three related tasks.The three groups of images are input to a network with AlexNet as the backbone,and feature sharing between tasks is performed through a shared unit.And under the effect of a series of learnable weights,learn the best sharing combination independently.Without any data enhancement,the method in this paper achieves an average accuracy rate of 96.111%,which is 6%higher than the traditional method,and the accuracy rate of the subclass is up to 27%.Experimental results show that the performance of defect detection networks based on multi-task feature layer sharing is better than similar single-task networks.