TY - JOUR
AU - Tang, Si
AB - The trade-off between detection speed and accuracy and the high hardware requirements of computing equipment have always been two major factors restricting the real-time detection and application of surface defects in aluminum strip. This paper proposes an effective, lightweight detection method for aluminum strip surface defects in industry, which improves the disadvantages of low efficiency and high calculation cost of the YOLOv4 framework. The backbone network GMANet is constructed based on a new convolution Ghost module, in which the union attention module is embedded in the stacked Ghost block. It realizes the compression of the network scale and focuses on the channel information of important feature maps. On this basis, the fusion neck network is redesigned and lightened by utilizing depthwise separable convolution and the sampling blocks of pixelshuffle and passthrough. It can reduce the information loss caused by sampling, and improve the extraction ability to multi-size features and the adaptive learning capability to weights. Moreover, the proposed method is trained and tested on the database of seven types of common defects collected from the quality inspection station of the cold rolling workshop. Experiments demonstrate that the proposed method achieves that the value of mAP is 94.68%, the model volume is reduced by 80.41%, and the detection speed is increased by three times, thereby outperforming the original YOLOv4 model. And it provides a research idea for the subsequent real-time detection of the aluminum strip surface on the embedded system.
TI - Automated real-time detection of surface defects in manufacturing processes of aluminum alloy strip using a lightweight network architecture
JF - Journal of Intelligent Manufacturing
DO - 10.1007/s10845-022-01930-3
DA - 2023-06-01
UR - https://www.deepdyve.com/lp/springer-journals/automated-real-time-detection-of-surface-defects-in-manufacturing-IoJ2QTVEpF
SP - 2431
EP - 2447
VL - 34
IS - 5
DP - DeepDyve
ER -