Abstract: Abstract: In order to solve the problems of non-compact models and low recognition accuracy and efficiency in current flame recognition algorithms, a lightweight flame image segmentation method based on salient target recognition theory is proposed.The method is based on a U-net-like encoder-decoder architecture, which uses salient target detection inside the architecture and introduces a multi-layer attention mechanism to detect flame targets in a hierarchical manner. The method achieves better recognition results on the public dataset. Comparing the four classical semantic segmentation models, it can be seen that the cross-comparison ratio index is improved by 5.70%~16.25%; the F1 score is improved by up to 10%; and the average absolute error value of the model in this paper is also much lower than the four classical models. It shows that the lightweight model in this paper has the best indexes in flame segmentation effect and operation speed, with strong robustness and effectiveness.

Key words: Key words: lightweight neural network, flame recognition, significant target recognition