Improved Flame Detection Algorithm Based on Salient Target Detection

Aiming at the problems of poor positioning accuracy, bad real-time performance, and potential safety hazards of shell positioning method in the range, a flame detection algorithm based on salient object detection is proposed. First, in view of missing datasets, a shell flame dataset is constructed for network model training and reasoning. Second, the parallel and crossed two-branch ResNet is used as the feature extraction module to learn foreground and background semantic information respectively. Furthermore, dilated convolution and attention mechanism are introduced to improve the receptive filed and synchronously enables the network to learn the ability of focusing on useful channels and spatial locations. Finally, the Bi-directional feature pyramid network (Bi- FPN) is introduced to fuse shallow texture information and deep semantic information to realize multi-scale and multi-stage prediction. Experimental results show that the proposed algorithm significantly outperforms the existing algorithms in terms of the accuracy, the regional integrity, and anti-interference, which is able to meet the needs of daily projectile positioning training in the shooting range.