基于轻量化卷积神经网络的陶瓷球表面缺陷快速识别方法  

Fast Recognition Method of Ceramic Ball Surface Defects Based on Lightweight Convolution Neural Network

作　　者：付鲁华[1] 庞家明 孙长库[1] 王鹏[1] FU Luhua;PANG Jiaming;SUN Changku;WANG Peng(State Key Laboratory of Precision Measuring Technology and Instruments,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学精密测试技术及仪器国家重点实验室,天津300072

出　　处：《传感技术学报》2023年第7期1041-1047,共7页Chinese Journal of Sensors and Actuators

摘　　要：由于球的形状特点,视觉方法需要处理多个角度的图像才能实现对单个球进行完整的缺陷识别,对图像处理速度要求较高。此外,卷积神经网络的浮点运算量(FLOPs)高,推理速度通常较慢。针对上述问题,基于MobileNetV3设计了更轻量化的卷积神经网络。首先通过改变宽度因子、减少基本单元数量、使用Ghost模块代替标准卷积降低原始网络参数量。最后通过坐标注意力机制提高网络对缺陷的识别准确率。实验结果表明,在氮化硅陶瓷球表面缺陷数据集上,提出的轻量化卷积神经网络相较于原始网络仅有2.2%的准确率损失。网络浮点运算量和参数量分别为原始网络的10.4%和3.3%,在边缘计算设备Jetson AGX Xavier上的推理时间小于7 ms,相较于原始网络提升超过40%,能够满足工业现场实时检测的需求。For vision method,multi angle images need to be processed to realize complete defect recognition of a single ball due to the shape characteristics,therefore,high image processing speed is required.In addition,convolutional neural network(CNN)has high float-ing-point operations(FLOPs)and the inference time is slow.To solve the above problems,a lighter convolutional neural network is de-signed based on MobileNetV3.Firstly,width factor is changed,basic unit’s numbers is reduced,and ghost module is used instead of standard convolution.The parameters and FLOPs were reduced by these methods.Finally,defects accuracy is improved through coordi-nate attention.The results show that the proposed lightweight convolution neural network has only 2.2%accuracy loss compared with the original.The floating-point operations and parameters of the modified network are 10.4%and 3.3%of those of the original CNN.The in-ference time on the edge computing device of Jetson AGX Xavier is less than 7 ms,which is more than 40%faster than the original CNN,meeting the needs of real-time recognition in industrial field.

关 键 词：缺陷识别 陶瓷球 GHOST 坐标注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]