Abstract: In order to solve the problems of large number of parameters for real-time flame detection and high requirements for hardware computing power, a lightweight real-time flame detection model based on improved YOLO v4-tiny is proposed. Firstly, the parameters of the model are pruned. Secondly, by adding improved Receptive Field Blocks (CSP-RFBs) in the shallow layer of the model, the perceptual field of the model shallow layer is improved. Thirdly, the framework of CSP-ResNet is improved, and the “hourglass CSP-ResNet” with faster and higher accuracy is proposed. Finally, a modified Spatial Pyramid Pooling (SPP) is adopted at the deep level of the model to further fuse the multiple sensory fields. The experimental results show that the accuracy of the improved YOLO v4-tiny model can reach 48.5%, which is 15.5% better than the original model. The number of parameters of the model  and the weight size of the weight file are 2.45BFLOPs and 16.3Mb, which are 63.9% and 30.6% less than the original model, respectively. The FPS on the mobile development board NVIDIA Jeston Xavier can reach 49.6, which is 21.9% better than the original model.

Key words: lightweight object detection, YOLO v4-tiny, real-time flame detection, receptive field, hourglass CSP-ResNet