Research on Flame Detection Based on Anchor-Free Algorithm FCOS

Currently, the deep learning anchor-based object detection algorithm has been widely used in flame detection. However, it relies too much on manual setting of anchor hyper-parameters and is insensitive to the change of object shape. Therefore, the improved anchor-free algorithm FCOS is introduced. Firstly, the Center-ness branch is replaced by the IoU prediction branch to make the bounding box location more accurate; then the random copy-pasting small objects and Mosaic data augmentation methods are used to improve the detection accuracy of small objects. The experimental results on the self-built dataset show that the AP and speed of FCOS are better than anchor-based algorithms. The use of IoU prediction branch can improve the location accuracy of the algorithm; the use of two data augmentation methods can further improve the AP of the algorithm for small objects by 9.6%.