Study on the Impact of Different Anchor Quantities in pneumonia X-ray Object Detection Based on the YOLO Series

While the COVID-19 pandemic may have ended, developing quick, efficient, and accurate diagnostic methods remains crucial. In this regard, analyzing chest X-ray images using machine learning technologies has demonstrated significant potential in the detection and diagnosis of COVID-19. This study delves into the influence of anchor quantities and different clustering algorithms in various object detection algorithms. Anchors act as reference boxes in the task of object detection in Convolutional Neural Networks (CNN), significantly impacting detection precision. Our research finds that optimizing the quantity of anchors can enhance detection performance by achieving a subtle balance to prevent overfitting. We evaluated different clustering algorithms like k-means and R-sample, each under scenarios with 3 and 9 anchors in the YOLO series algorithms, on their impact and performance on mAP, LAMR, Recall, Precision, Fl Score, and AP. Additionally, we compared these with FRCNN and anchor-free models like YOLOX and Y0L0v8. We concluded that adjusting the number of anchors and applying specific clustering algorithms’ synergistic effects under the same dataset, certain combinations can offer superior performance in speed and accuracy of COVID-19 detection than current mature anchor-free detection techniques.