精密刀具表面缺陷认知神经网络模型研究

Research on Surface Defect Cognition Neural Network Model for Precision Cutter Tool

摘要:

精密刀具表面凸起齿的正刀口在使用过程中容易产生磨损和磕碰缺陷，这种缺陷存在尺度小、纹理多样等特点，传统机器学习模型难以实现高精度检测。为此本文提出一种基于改进You Only Look Once Version 8 (YOLO-v8)的刀具缺陷检测方法（CutterNet）。首先，提出Cross-Stage Partial Multi-Scale Attention Module (CSPMSAM)，用于提取刀具局部缺陷特征，增强模型对不同尺度缺陷的检测能力；其次，引入Asymptotic Feature Pyramid Network (AFPN)，加强不同尺度特征间融合，缩小它们之间的信息差距；最后，使用Inner-CIoU( Complete Intersection over Union )替换YOLO-v8中CIoU损失函数，增强边界框的回归结果。实验结果表明，改进后的算法在检测准确率方面提升了3.1%，模型参数量下降30.12%，推理速度由58帧提升到60帧，优于其他大多数主流目标检测模型，该算法已经应用到刀具缺陷实时检测系统。

Abstract:

Surface of cutter front side with raised teeth is prone to be worn and bumped under high-frequency using. These defects are characterized by small scales and diverse shapes, making it difficult for traditional machine learning models to achieve high-precision. To address this issue, a cutter defect detection method (CutterNet) was proposed based on an improved YOLO-v8. First, the Cross-Stage Partial Multi-Scale Attention Module is introduced (CSPMSAM) to extract local features and enhance the model's ability solving different scales. Secondly, the Asymptotic Feature Pyramid Network (AFPN) is incorporated to strengthen the fusion of features across different scales and reduce the information gap between them. Finally, the CIoU loss function in YOLO-v8 with the Inner-CIoU (Complete Intersection over Union) is replaced to improve the regression results of the bounding boxes. Experimental results indicate that the improved algorithm increases detection accuracy by 3.1%, reducing the model parameter amount by 30.12%. The inference speed is also improved from 58 to 60 frames. This performance surpasses the up-to-date object detection models. This algorithm has been applied to a real-time detection system for tools defects.