Dba-Yolo: A Lightweight Fire Image Detection Algorithm with Structural Reparameterized Multiple Attention Mechanisms

Abstract

In response to the challenges of detecting smoke-like objects in fire scenes and fire detection under various natural lighting conditions, we propose an improved fire detection algorithm based on YOLOv7-Tiny, named DBA-YOLO. This novel approach introduces an attention feature extraction module that combines structural re-parameterization, extending the feature extraction network architecture. It enhances feature propagation for flame target recognition and improves the performance of the network model. The integration of a focal modulation network into the head detection layer facilitates better feature exchange and enlarges the receptive field of the detection layer. By replacing the original model's detection head with a variable attention detection head, we significantly increase detection precision. The synergistic effect of these three attention modules within the network further enhances detection capabilities. Experimental results demonstrate that our algorithm, especially in detecting fire and smoke targets, achieves high detection accuracy and low computational cost compared to the current state-of-the-art object detection networks. It attains an average detection precision of 85.4% and a mean Average Precision (mAP50) of 85.6%. Compared to the popular baseline algorithm YOLOv7, our proposed DBA-YOLO algorithm achieves a 3.5% increase in mAP and a 2.1% increase in accuracy. With an input image size of 640x640 pixels, the average detection time per frame is 6.5ms, which is 27% faster than the YOLOv7 model, surpassing most current target detection models in all fire detection metrics including accuracy, recall, and mean precision. The proposed model effectively detects fire targets as well as fire-like and smoke-like objects, minimizing false positives with high accuracy and is applicable to target detection in other complex scenarios.

Keywords: Fire Detection, Structural Reparameterization, Multiple Attention, Smoke Detection, Convolutional Neural Network