Flame and Smoke Detection Algorithm Based on ODConvBS-YOLOv5s

Real-time and accurate detection of flame and smoke is an important prerequisite to reduce the loss caused by fire. There are exists some problems in traditional detection algorithms of flame and smoke, such as low accuracy, high miss rate, low detection efficiency, and low detection rate of small targets. This paper proposes the detection algorithm of flame and smoke based on ODConvBS in YOLOv5s. Firstly, the ordinary convolutional blocks in the backbone network of YOLOv5s are replaced with ODConvBS to achieve the extraction of attentional features from the convolutional kernel. Secondly, Gnconv is introduced into Neck to improve the high-order spatial information extraction ability of the model. Then, the Shuffle Attention module is added at the end of the Neck to facilitate the fusion of different groups of features. Finally, the prediction part uses a SIOU loss function that can take into account the angle of the prediction frame vectors to accelerate the model convergence. When utilizing the self-made dataset of flame and smoke, the upgraded YOLOv5s model mAP grew by 9.3%. At the same time, the accuracy rate, the recall rate, and the detection speed increased to 83.5%, 83.7%, 33.3FPS respectively.