Flame image recognition faces a large model size, low accuracy, and poor real-time performance in edge equipment and mobile terminal equipment environments. In order to solve these problems, firstly, ShuffleNetV2 is selected as a lightweight backbone neural network to ensure the real-time performance of the model. Secondly, a new space and channel dual attention module (SCDAM) is designed to analyze the relevance of channels and spaces, and different weights are given according to the importance of different features, so as to effectively improve the model accuracy. Then, in order to enrich the extracted features on the spatial scale, a multi-scale feature fusion module is designed to enhance the adaptability of the network to different scales. Finally, the SCDAM and multi-scale module are introduced into ShuffleNetV2, and the model parameters are optimized by transfer learning, so as to further improve the model accuracy. With only a slight increase in the amount of parameters and calculations, the accuracy of the proposed algorithm is 3.2% higher than that of ShuffleNetV2, and the single inference takes only 8.7 ms. Experiments show that the proposed algorithm is more suitable for conditions with limited computing resources, such as the identification and monitoring of gunpowder flame.