Improved flame detection based on global attention mechanism and deformable convolution

While convolutional neural networks (CNNs) like YOLOv8 have advanced object detection, the original YOLOv8 exhibits limitations in fire scenario applications: insuffi- cient sampling of fixed rectangular structures and inability to accurately recognize small flame targets.Thus, the study aims to address these limitations by improving YOLOv8, enabling more effective real-time monitoring and accurate identification of fire and smoke- ultimately enhancing the practicality of fire detection technologies in complex real-world scenarios. This paper conducts a more thorough research and discussion on the detection of flame and smoke objects. To achieve real-time and accurate monitoring of smoke and fire targets in fire scenarios, the Improve-YOLOv8 algorithm is proposed based on YOLOv8. Firstly, this model replaces the C2f module with the DCNv2 module at the Backbone end, which can adaptively adjust according to the proportion and morphology of the flame and smoke targets. This design effectively addresses the issue of insufficient sampling of the fixed rectangular structure, thereby enhancing the models accuracy. Secondly, a small object detection layer is added in the Head part to address the challenge of accurately recognizing small flame targets. Additionally, the GAM attention mechanism is integrated after the Backbone SPPF module to further improve the accuracy and focus of image processing, thereby enhancing the overall processing efficiency. Finally, migration learning is implemented using sample data to effectively differentiate flame and smoke information and enhance detection accuracy. The experimental results show that the accuracy of the improved algorithm model reaches 98.6%, and the accuracy is improved by 10.5% compared with the original YOLOv8n, and m A P @ 0 . 5 9.6%. The comprehensive performance is improved, and the model can realize the real-time monitoring and accurate identification of fire and smoke.