Zhang, Jiaqiao Tokushima University|Nantong University
Kang, Xin Tokushima University Tokushima University Educator and Researcher Directory
Ni, Hongjun Nantong University
Ren, Fuji Tokushima University Tokushima University Educator and Researcher Directory KAKEN Search Researchers
The steel strip is one of the essential raw materials in the machinery industry. Besides, the defects on the surface of the steel strip directly determine its performance. To achieve rapid and effective detection of surface defects on steel strips, a CP-YOLOv3-dense (classification priority YOLOv3 DenseNet) deep convolutional neural network was proposed in the present study. The model used YOLOv3 as the basic network, implemented priority classification on the target images, and then replaced the two residual network modules in the YOLOv3 network with two dense network modules. Therefore, the model can receive the multi-layer convolution features output by the dense connection block before making predictions, consequently enhancing the reuse and fusion of features. Finally, the six kinds of surface defects of steel strips were detected by the improved network model, and the results were compared with other deep learning networks. According to the results, the recognition precision of the CP-YOLOv3-dense network model is 85.7%, the recall rate is 82.3%, the mean average precision is 82.73%, and the detection time of each image is 9.68ms. The mean average precision is 6.65% higher than the original YOLO network and 10.6% higher than the DNN network. In addition, the detection speed is 1.77 times faster than the Faster RCNN network. The proposed CP-YOLOv3-dense network has stronger robustness and higher detection precision, which can be used for the identification of various steel strip surface defects.
Ironmaking & Steelmaking
Institute of Materials, Minerals and Mining|Taylor & Francis
This is an Accepted Manuscript of an article published by Taylor & Francis in Ironmaking & Steelmaking on 15/09/2020, available online: http://www.tandfonline.com/10.1080/03019233.2020.1816806.
|DOI (Published Version)|
|URL ( Publisher's Version )|
Science and Technology