The deepening of the number of network layers can weaken the ability to characterize the detailed information of the deep features of the flame target, and at the same time extract redundant features with low correlation, resulting in low flame recognition accuracy. Aiming at this problem, a flame detection method based on improved Faster R-CNN is proposed to improve the accuracy of flame recognition in deep networks. Firstly, the ResNet50 network is used to extract flame features, and the SENet module is added to reduce the redundant features of flame targets. Then, the deep features and shallow features are multi-scale feature fusion to enhance the detailed information of deep features. Finally, the network is trained to realize the recognition of flame targets positioning. In the experiment, the VOC flame data set is constructed for network training, the test set is used for detection, and the feature map visualization is compared. Compared with the model before the improvement, the AP value increases by 7.78%, the recall increases by 9.05%, and the precision increases by 12.54%. By combining the attention mechanism module and the multi-scale feature fusion mechanism, the flame target detection model proposed in this paper, can effectively extract the flame target feature, and the flame target detection result is more accurate.