计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (22): 262-270.DOI: 10.3778/j.issn.1002-8331.2201-0218

• 工程与应用 • 上一篇    下一篇

改进FSSD的差速器壳体表面缺陷实时检测算法

王国东,唐金亮,陈特欢,崔杰   

  1. 1.宁波大学 机械工程与力学学院,浙江 宁波 315211
    2.浙江华朔科技股份有限公司,浙江 宁波 315800
  • 出版日期:2022-11-15 发布日期:2022-11-15

Real-Time Detection Algorithm for Differential Housing Surface Defects Based on Improved FSSD

WANG Guodong, TANG Jinliang, CHEN Tehuan, CUI Jie   

  1. 1.College of Mechanical Engineering and Mechanics, Ningbo University, Ningbo, Zhejiang 315211, China
    2.Zhejiang Huashuo Technology Co., Ltd., Ningbo, Zhejiang 315800, China
  • Online:2022-11-15 Published:2022-11-15

摘要: 针对传统检测方法对于汽车差速器壳体表面小目标缺陷的误检和漏检问题,提出了一种改进的FSSD_MobileNet缺陷检测模型。该模型将FSSD(feature fusion single shot multibox detector)算法的基础骨干网络VGG16替换成轻量级MobileNet网络,构建了一种高效的特征融合结构并调整了默认框的尺寸,进一步提升对小目标缺陷的检测能力。同时使用RMSProp(root mean square propagate)梯度下降算法来优化损失函数,加快了模型的收敛速度。实验结果表明,改进后的FSSD_MobileNet模型的mAP为96.7%,相比于改进前提升了16.2个百分点。在保持较高检测精度的同时,检测速度达到了191?FPS,高于目前单阶段算法中速度较快的YOLOv5s网络,相较于传统的SSD(single shot multibox detector)和FSSD分别提升了94?FPS和102?FPS,同时模型较为精简,能够更好地满足实际生产中对准确性和实时性的综合要求。

关键词: 目标检测, 特征融合, FSSD, 深度可分离卷积, 小目标缺陷

Abstract: Aiming at the traditional detection methods for the problems of false detection and missed detection of small target defects on the surface of automobile differential housing, an improved FSSD_MobileNet defect detection model is proposed. The model replaces the basic backbone network VGG16 of the FSSD(feature fusion single shot multibox detector) algorithm with the lightweight MobileNet network, constructs an efficient feature fusion structure. The model also adjusts the size of the default box, which further improves the detection ability of small target defects. Meanwhile, the RMSProp(root mean square propagate) gradient descent algorithm is used to optimize the loss function, which speeds up the convergence speed of the model. The experimental results show that the mAP of the improved FSSD_MobileNet model is 96.7%, which is 16.2 percentage points higher than that before the improvement. While maintaining high detection accuracy, the detection speed has reached 191 FPS, which is higher than the faster YOLOv5s network in the current one-stage algorithm. Compared with the traditional SSD(single shot multibox detector) and FSSD, the detection speed has increased by 94 FPS and 102 FPS respectively. At the same time, the model is relatively simplified, which can better meet accuracy and real-time requirements comprehensively in actual production.

Key words: target detection, feature fusion, FSSD, depthwise separable convolution, small target defects