Anchor-Free Smoke and Flame Recognition Algorithm with Multi-Loss

Abstract

Fire perception based on machine vision is essential for improving social safety. Object recognition based on deep learning has become the mainstream smoke and flame recognition method. However, the existing anchor-based smoke and flame recognition algorithms are not accurate enough for localization due to the irregular shapes, unclear contours, and large-scale changes in smoke and flames. For this problem, we propose a new anchor-free smoke and flame recognition algorithm, which improves the object detection network in two dimensions. First, we propose a channel attention path aggregation network (CAPAN), which forces the network to focus on the channel features with foreground information. Second, we propose a multi-loss function. The classification loss, the regression loss, the distribution focal loss (DFL), and the loss for the centerness branch are fused to enable the network to learn a more accurate distribution for the locations of the bounding boxes. Our method attains a promising performance compared with the state-of-the-art object detectors; the recognition accuracy improves by 5% for the mAP, 8.3% for the flame AP50, and 2.1% for the smoke AP50 compared with the baseline model. Overall, the algorithm proposed in this paper significantly improves the accuracy of the object detection network in the smoke and flame recognition scenario and can provide real-time fire recognition.