LE-YOLOv5: A Lightweight and Efficient Neural Network for Steel Surface Defect Detection

被引：0
作者：
Zhu, Chengshun [1 ]
Sun, Yong [1 ]
Zhang, Hongji [1 ]
Yuan, Shilong [1 ]
Zhang, Hui [1 ]
机构：
[1] Jiangsu Univ Sci & Technol, Sch Mech Engn, Zhenjiang 212100, Peoples R China
来源：
IEEE ACCESS | 2024年 / 12卷
关键词：
Convolution; Computational modeling; Feature extraction; Steel; Defect detection; Surface treatment; Detectors; Accuracy; Transformers; Neural networks; Deep learning; YOLOv5; surface defect detection; attention mechanism;
D O I：
10.1109/ACCESS.2024.3519161
中图分类号：
TP [自动化技术、计算机技术];
学科分类号：
0812 ;
摘要：
Due to the influence of manufacturing process and external factors, there will be some undesired defects on the steel surface, which seriously affects the lifetime of steel, and the traditional surface defect detection efficiency and speed are not satisfactory. Therefore, based on the industrial scenario of low computational force, this study proposed a lightweight and efficient defect detector called LE-YOLOv5. First, we utilize ShuffleNetv2 as the backbone of the model, which greatly reduces the number of parameters. Second, we propose a CBMM module to expand the global receptive field of the model in the initial down sampling stage, which facilitates the model in capturing global information. Third, we also propose a parallelized C2N module for the detection of small defects. Finally, we design a global coordination attention (GCA) to efficiently connect position and spatial information from the feature map. Numerous experimental results demonstrate that LE-YOLOv5 has a highly superior overall performance, reaching 79.1% mean Average Precision (mAP) on the NEU-DET dataset while inferring an image on the CPU in 196.1 ms, which is 5% and 1.5% improved mAP compared to YOLOv5M and YOLOv5L, respectively. At the same time, under the condition that the inference time for an image on a CPU-dependent low computing power force remains the same, the accuracy has improved by 5.3% compared to YOLOv8. It provides excellent potential for defect detection of steel in industrial environment.
引用
收藏
页码：195242 / 195255
页数：14
相关论文
共 56 条
  • [31] Yuan L., Chen Y., Wang T., Yu W., Shi Y., Jiang Z., Tay F.E.H., Feng J., Yan S., Tokens-to-Token ViT: Training vision transformers from scratch on ImageNet, Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 538-547, (2021)
  • [32] Tian Z., Shen C., Chen H., He T., FCOS: Fully convolutional onestage object detection, Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 9626-9635, (2019)
  • [33] Wang Y., Zhang X., Yang T., Sun J., Anchor DETR: Query design for transformer-based detector, Proc. AAAI Conf. Artif. Intell., 36, 3, pp. 2567-2575, (2022)
  • [34] Xiao A., Transformer in Transformer, (2021)
  • [35] Howard A., Sandler M., Chen B., Wang W., Chen L.-C., Tan M., Chu G., Vasudevan V., Zhu Y., Pang R., Adam H., Le Q., Searching for mobilenetv3, Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 1314-1324, (2019)
  • [36] Howard A.G., Zhu M., Chen B., Kalenichenko D., Wang W., Weyand T., Andreetto M., Adam H., MobileNets: Efficient convolutional neural networks for mobile vision applications, (2017)
  • [37] Sandler M., Howard A., Zhu M., Zhmoginov A., Chen L.-C., MobileNetV2: Inverted residuals and linear bottlenecks, (2018)
  • [38] Ma N., Zhang X., Zheng H., Sun J., ShuffleNet V2: Practical guidelines for efficient CNN architecture design, Proc. ECCV, pp. 122-138, (2018)
  • [39] Zhang X., Zhou X., Lin M., Sun J., ShuffleNet: An extremely efficient convolutional neural network for mobile devices, (2017)
  • [40] Iandola F.N., Han S., Moskewicz M.W., Ashraf K., Dally W.J., Keutzer K., SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0. 5MB model size, (2016)
123456