A A NOVELTY METHOD FOR PCB DEFECT DETECTION ON YOLOv8 BASIS

Abstract

To tackle the challenges in PCB defect detection, I’ve developed a new algorithm built on an improved YOLOv8 framework. This approach is aimed at boosting detection accuracy while reducing model complexity, making it well-suited for detecting smaller targets and functioning effectively in environments with limited resources. The algorithm starts by introducing a refined neck network structure, which cuts down on the number of model parameters and computational demands, improving how efficiently resources are used. Additionally, the inclusion of ShuffleAttention and a BiFPN structure strengthens the model's ability to fuse features at multiple scales, significantly enhancing its performance with smaller targets.

On top of that, I’ve replaced the commonly used CIoU loss function with a WIoU loss function, which makes the model more accurate and robust in its detection capabilities. In tests, this enhanced model achieved impressive results, with mAP50 and mAP90-95 scores reaching 93.4% and 48.3%, respectively. What's more, model parameters, GFLOPs, and weight size were reduced by 33%, 12%, and 32%, respectively, bringing them down to 1.882M, 7.0, and 4.3M. This makes the solution not only highly efficient and accurate but also lightweight—perfect for use in constrained environments like mobile devices and embedded systems.