Target Defect Detection Based on YOLOv5

Abstract

This paper introduces a defect detection system based on YOLOv5 and K-Means clustering algorithm, aimed at efficiently and accurately detecting surface flaws in targets. In the context of high costs, complex operations, and stringent environmental requirements posed by traditional flaw detection methods, this system integrates deep learning technology and an optimized K-Means algorithm, significantly enhancing the efficiency and accuracy of the YOLOv5 model in steel flaw detection. Innovations to the original backbone network of YOLOv5 include the addition of multi-layer upsampling, effectively improving the detection capability for flaws of various scales. Furthermore, by augmenting the target flaw dataset, this system not only enriches the sample volume but also strengthens feature extraction capabilities by integrating novel convolutional structures and an improved attention mechanism. Experimental results demonstrate significant improvements in average precision and recall rates for the modified YOLOv5 detection model on the flaw dataset, achieving 70% of the original model's detection speed and fully meeting the requirements for flaw detection in industrial production settings.