An efficient and lightweight image detection model based on YOLOv8 and its application

An efficient and lightweight image detection model based on YOLOv8 is presented, and it is applied to detect the surface defects on steel strips in this paper. Under the framework of YOLOv8, a multi-scale detail enhancement mechanism is firstly inserted to ensure the robustness of multi-scale targets detection, and a shared convolution is used to enhance the detail-sharing capability between different detection layers. A LSDECD is designed to improve the detection accuracy of the presented model for multi-scale targets, especially small targets and the inference capabilities in complex tasks. A Cross Stage Partial with Two Conv Blocks fused (C2f) - Star-Context-Aware Anchor Attention (CAA) module is designed to enhance feature extraction capabilities. The Backbone of YOLOv8 is replaced with the StarNet network, which can significantly reduce the computational complexity and minimize the number of parameters. The presented model is trained and verified by using the surface images of hot-rolled steel strips (HRSSs) dataset and GC10-DET dataset. The experiment results indicate that the mAP reaches 82.1% and 72.1%, respectively and the complexity and parameter numbers of the model reduce 25% and 28%, respectively.