Improved YOLOv5 Based on Feature Fusion and Attention Mechanism and Its Application in Continuous Casting Slab Detection
Hongliang Yang, Yiming Fang, Le Liu, Hao Ju, Kesong Kang
Abstract
Object detectors based on convolutional neural network (CNN) have been widely used in industrial production. In the production and transportation process of continuous casting slab, it is necessary to dynamically track the slab logistics and orderly process the tracking information when the slab is ON or OFF the production line. In this article, the continuous casting slab dataset on the continuous casting slab transfer roller table is first established, which includes photos of slabs, crown blocks, and crown blocks holding slabs under various lighting conditions on the continuous casting roller table. Then, an improved you only look once version 5 (YOLOv5) frame called VCIoU, ACON-C, attention networks, and BiFPN (VAAB)-YOLOv5 [YOLOv5 with vertex center intersection over union (VCIoU), the attention network, activate or not (ACON)-C activate function, and bidirectional feature pyramid network (BiFPN)] is proposed for the detection of continuous casting slabs. VAAB-YOLOv5 uses a new loss called VCIoU loss, which introduces the feature of distance between the vertex and center point. The VCIoU loss improves the detection accuracy without increasing the calculation cost. In addition, this framework also improves the detection accuracy by improving the network structure and activation function, including adding the attention mechanism, introducing the ACON-C activation function, and improving FPN. At last, the experimental results show that the mean average precision (mAP)@50:95 index of the proposed method on the self-established continuous casting slab dataset is 87.75%, which is 1.58% higher than the YOLOv5 small (YOLOv5s) algorithm. The VAAB-YOLOv5 improves the detection accuracy while meeting the speed requirements of slab detection on the continuous casting site.