ZFD-Net: Zinc flower defect detection model of galvanized steel surface based on improved YOLOV5

Abstract

Due to the complex factory environment, zinc flower defects and galvanized sheet background are difficult to distinguish, and the production line running speed is fast, the existing detection methods are difficult to meet the needs of real-time detection in terms of accuracy and speed. We propose ZFD-Net, a zinc flower defect detection model on the surface of galvanized sheet based on improved you only look once (YOLO)v5. Firstly, the model combined the YOLOV5 model with our proposed cross stage partial transformer (CSTR) module in this paper to increase the model receptive field and improve the global feature extraction (FE) capability. Secondly, we use bi-directional feature pyramid network (Bi-FPN) weighted bidirectional feature pyramid network to fuse defect details of different levels and scales to improve them. Then we propose a cross resnet simam fasternet (CRSFN) module to improve the reasoning speed of ZFD-Net and ensure the detection effect of zinc flower defects. Finally, we construct a high-quality dataset of zinc flower defect (ZFD) detection on galvanized sheet surface, which solves the problem that no public dataset is available at present. ZFD-Net is compared with state-of-the-art (SOTA) methods on the self-built data set, and its performance indicators are better than all methods.

Fig 1. Picture of sample zinc-coated steel.

Fig 2. Pictures from the factory production line.

Fig 3. Examples of different defects.

Fig 4. Reduce motion blur.

Fig 5. Effect images of CLAHE algorithm.

Fig 6. Effect images of data enhancement.

Fig 7. Overall structure of ZFD-Net.

Fig 8. CSTR module structure diagram.

Fig 9. Several features of the pyramid structure.

Fig 10. CRSFN module structure diagram.

Fig 11. Images of test set.

Fig 12. ZFD-Net defects effect.

Fig 13. Several schemes to detect twill defects effect.

Fig 14. Several schemes to detect inequality defects effect.