Lightweight Target Detection: An Improved YOLOv8 for Small Target Defect Detection on Printed Circuit Boards

Addressing the performance and efficiency challenges of detecting small target defects on PCB (printed circuit board), this research proposes a lightweight inspection model for YOLOv8-based. The initial step involves employing an enhanced Efficient_Detect network architecture with the aim of diminishing the volume of model parameters and computational intricacy, while simultaneously augmenting resource utilization. Then, the deformable convolutional networks v2 (DCNv2) are integrated into the C2f structure to improve feature extraction for defects with diverse shapes on the PCB surface, enhancing the model's capability to identify different categories of flaws. Meanwhile, the Inner-SIoU loss function is utilized instead of the conventional CIoU loss function to enhance both the precision of inspection and the robustness of the model. The enhanced approach achieves an mAP50 of 96.8% and an mAP50-95 of 72.2%, marking an improvement of 3.9% and 10.1%, respectively, comparison with base YOLOv8n model. In addition, the parameter counts of the model, GFLOPs, and weight size are notably reduced by 18.4%, 33.3%, and 19.7%, respectively, resulting in final metrics of 2.34 M, 5.4 GFLOPs, and a weight size of 4.9 M. In addition, the detection speed can reach up to 161 FPS, which has the potential for real-time detection. It provides an efficient and lightweight PCB small target defect detection solution for scenarios with limited computing resources.