Algorithm for steel defect detection based on YOLOv7-tiny

Steel is widely used in many fields such as construction, automobile, machinery, bridge, ship and so on in China. Steel surface defects can seriously affect the service life and performance of steel. To tackle the issues related to steel surface defect detection process exists in the detection algorithm parameter quantity is large, difficult to transplant to small terminal equipment, defect target leakage detection and other problems. This paper offers an optimized YOLOv7-tiny detection algorithm through improvements. In the following paper, the initial modification involves upgrading the up-sampling segment of the Neck in the YOLOv7-tiny design, and replace the recent interpolation in the up-sampling part with the lightweight operator CARAFE, to overcome the challenge posed by the dense distribution of defects on the steel surface, making detection difficult. Secondly, the triple coordinate attention mechanism is utilized to augment the network model's feature extraction effectiveness, so as to improve the detection accuracy. Thirdly, GSConv is integrated into the ELAN-T submodule of the YOLOv7-tiny network architecture, and the ELAN-G module is implemented to streamline the model's parameters and reduce computational overhead, all while improving accuracy, so as to realize the lightweight network structure.