Abstract: The application of image processing technology based on various convolution neural network algorithms to detect and identify surface defects can not only reduce the cost of labor,but also greatly improve the efficiency and accuracy.However,the current popular image processing technology has the characteristics of large computation,high storage cost and very complex,which is contrary to the high real-time and limited computing resources required by industrial applications.Therefore,a Multi-scale Compression Convolution Neural Network model (MC-CNN) was proposed for the rapid detection of steel surface defects.The multi-scale compression of the network was carried out by network structure optimization,knowledge distillation,network pruning and parameter quantization.The experimental results showed that the proposed method could greatly improve the recognition efficiency,reduce the volume of the model,which was facilitate the application in various scenarios with high real-time requirements and limited storage and computing resources.

Key words: surface defects detection, convolution neural network, multi-scale, network compression

摘要： 应用基于各种卷积神经网络算法的图像处理技术对表面缺陷进行检测识别,不仅可以降低人工的成本,还可以大大提高效率和准确度。但是,当前比较热门的图像处理技术普遍存在计算量大、存储成本高以及模型复杂等特点,与工业应用所要求的高实时性以及有限的计算资源相悖。因此,提出一种基于多尺度压缩卷积神经网络模型(MC-CNN)进行表面缺陷快速检测,通过网络结构优化、知识蒸馏、网络修剪以及参数量化等压缩方法对网络进行多尺度压缩。实验表明,所提方法能够在保证卷积神经网络识别准确度的同时,大幅度提高模型的识别效率,缩小模型的体积,便于在工业现场各种实时性要求高而且存储和计算资源有限的场景中应用。

关键词: 表面缺陷检测, 卷积神经网络, 多尺度, 网络压缩