An Improved YOLOv5s Fire Detection Model

It is well-established that contact fire sensors are susceptible to interference from non-fire particles and cannot be applied to fire alarms in both large indoor and outdoor open spaces. On the other hand, the fire detection technology based on the image has several advantages including non-contact, fast response, strong anti-interference, unlimited application space, and comprehensive fire alarm information, which is preferable to fire detection. To this end, in this work, eight existing object detection models were first compared based on convolutional neural networks by autonomously collecting flame smoke datasets. From the acquired results, it was demonstrated that the YOLOv5 network has higher Mean Average Precision (mAP) and Frame Per Second (FPS) than the others. Next, further optimization of the YOLOv5s network was carried out. By introducing Convolutional Block Attention Module (CBAM), replacing PANet with BiFPN, and replacing nearest neighbor interpolation with transposed convolution (TC), the accuracy of the YOLOv5s network was significantly improved. Simultaneously, three lightweight networks, namely MobileNetV3, ShuffleNetV2, and GhostNet, were used to lighten the YOLOv5s network. The validation results indicated that the mAP of the improved YOLOv5s model reached 82.1%, the parameters were as low as 5.9 M and the floating-point operations (FLOPs) dropped to 8.1G. Finally, Single Eye (SiEYE) and Double Eyes (DoEYE) detection networks were proposed for emergency conditions based on the above-mentioned improvements. From the test results, it was indicated that the network has stronger robustness and meets the fire detection requirements, which is of great importance for the next research on smoke diffusion prediction.