Flame Detection using YOLOv5 with K-means Optimized Anchor Boxes

Object detection is a challenging task in the domain of computer vision, aiming to identify and localize objects within images or videos. However, there are still research gaps in enhancing detection accuracy and processing speed, especially in complex backgrounds and with various object scales. In this study, we adopt an improved YOLOv5 model incorporating K-means clustering to optimize the anchor boxes, aiming to enhance flame detection performance. The proposed approach features the integration of K-means for better anchor box initialization, which enables more accurate bounding box predictions. Through comparative experiments on a self-constructed dataset, the results demonstrate that the K-means optimization strategy effectively improves the detection capability of YOLOv5. The accuracy of the proposed model increased from 86.91% to 88.29%, confirming the effectiveness of the optimization.