In recent years, with the development of infrared sensor technology, many image processing applications based on infrared images have emerged. Infrared images can provide 24/7 information that the human eye can't see. Therefore, infrared image is widely used in monitoring, industry, military and other fields. Unlike visible light images captured by ordinary sensors, infrared images typically have low contrast, contain blurred edges, and a lot of noise. The reason for the low contrast and blurred edges is that usually the foreground and background have similar temperatures. Low contrast and blurred edges will produce low-quality infrared images. In addition, infrared images have a low signal-to-noise ratio due to the infrared sensor and readout circuit of infrared cameras, which leads to low signal and high noise problems that further degrade infrared image quality. And low-quality infrared images can bring many difficulties to further analysis, such as object recognition and image fusion. It is best to use effective infrared image enhancement techniques to produce high-quality infrared images, that is, high contrast, clear detail and less noise.In this paper, an infrared image enhancement algorithm based on multi-scale feature extraction and fusion is proposed. Firstly, a multi-scale adaptive feature extraction fusion module is designed, which fuses convolution features detected at multiple levels based on multi-layer feature fusion to fully preserve details in image reconstruction. The Global Attention Mechanism (GAM) is introduced into the multi-scale adaptive feature extraction fusion module, which can amplify the interdimensional interaction, obtain the features of three dimensions at the same time, avoid information loss, and retain more feature information. Then, the luminance enhancement iteration function is designed, using the fusion features of different levels as the pixel parameters of the iteration function to avoid the problem of local exposure. Finally, a feature fusion and image reconstruction module is designed. After two convolution with residual structures, convolution with different expansion rates is used to form a double-branch structure, which can enhance feature propagation. Finally, the cross-channel information fusion is realized by using two convolution. This paper outputs the characteristics of each convolution layer in 6 stages.Based on the above methods, two different data sets were first selected for subjective comparison (Fig.10, Fig.11) and objective index evaluation (Tab.1, Tab.2) between the proposed algorithm and five other popular image enhancement algorithms. The results showed that whether the text algorithm was applied to self-built data sets or MSRS public data sets. The brightness enhancement is moderate and is almost the best among all algorithms in texture feature retention, indicating that the proposed algorithm has good performance and high applicability in feature integrity preservation and brightness enhancement. Then, the multi-scale network structure (Tab.3); The overall aspect of the algorithm in this paper (Tab.4); The results of the ablation experiments on the selection of different dilatation convolution rates (Tab.5) and the selection of parameters of the luminance enhancement iteration function (Tab.6) show that the proposed algorithm performs well in various improvement choices. Finally, in order to ensure the application effectiveness of the image enhancement algorithm in this paper, the STS (Single Shot MultiBox Detector-Threshold Segmentation) fire hazard detection algorithm is introduced. The proposed algorithm was compared with the images enhanced by other five algorithms for fire hazard detection (Fig.13, Tab.7). According to the results, the accuracy rate of early fire detection of the proposed algorithm was 97.86%, the accuracy rate of hidden hazards and non-hidden hazards in the images were 97.13% and 98.59%, respectively, and the false detection rate was 1.02%. Compared with the five algorithms in the above literatures, the results increased by 9.11, 5.96, 16.31, 4.95 and 4.15 percentage points respectively. Moreover, the above evaluation indicators are consistent with the results of the subjective evaluation method in Fig.10, which proves that the algorithm has a good effect on early fire detection.In this paper, an infrared image enhancement algorithm with multi-scale feature extraction and fusion is proposed. Firstly, multi-scale adaptive feature extraction module with global attention mechanism is introduced to extract multi-scale features from input images, and multi-channel Concat is performed to obtain finer texture features. Then, the luminance enhancement iteration function using fusion feature graph as parameter is used to generate global enhancement through multiple iterations. Finally, the proposed feature fusion and image reconstruction module fuses various features to obtain enhanced infrared images. Theoretical analysis and experimental results show that the proposed algorithm is more competitive than the more mature enhancement methods in recent years on both self-built and MSRS public data sets. And the enhanced early fire warning algorithm application experiment is carried out, which shows the effectiveness and feasibility of the infrared image enhancement algorithm of multi-scale feature extraction and fusion in this paper.