In order to realize laboratory fire recognition and solve the problems that the flame is not significant in the image collected by the camera due to the small fire, and the flame with smoke occlusion affects the accuracy of segmentation and recognition, an improved semantic aware real-time thermal infrared and visible image fusion segmentation network is proposed. The thermal radiation information is provided to enhance the spectral information reduced by smoke occlusion in the visible light image, as well as the significance of the flame in the early stage of combustion, and the segmentation of the flame under the laboratory smoke occlusion and the small flame in the early stage of flame combustion is completed. The intermediate feature transfer block (IFTB) is added to the gradient residual dense block (GRDB) in the fusion network, and the weight block is introduced to reduce the information loss of the flame image during fusion, and the visual image structure information is restored with the minimum content loss as the benchmark while enhancing the saliency of the flame image. The edge extraction module based on gradient transformation (EEM) is added to the Deeplabv3+ semantic segmentation network to enhance the edge information of flame and smoke images with significant light and dark transformation in the fusion image, reduce the influence of smoke occlusion on flame segmentation, and improve the accuracy of flame segmentation and recognition. The experimental results show that the accuracy of flame detection segmentation and recognition in the early stage of flame combustion is improved by fusing visible light and thermal infrared images. The average intersection over union ratio of the improved flame segmentation network in the self-collected data set is 91.27%, and the segmentation efficiency is 11.96 FPS. The improved fusion segmentation network significantly improves the effect of laboratory flame and smoke segmentation and recognition, and has practical application value for laboratory flame and smoke detection.