【摘要】 有害气体泄漏检测在工业安全、环境保护以及应急响应中至关重要。现有方法普遍基于可见光与红外双波段成像系统，但传统融合方法在双波段系统中直接应用时，往往导致关键气云目标信息弱化，从而影响检测精度与可靠性。为此，提出一种基于时空域显著性的双波段气体图像融合增强方法。该方法首先基于红外序列图像提取时域中的气云目标，然后利用多尺度滤波分解可见光、红外和时域目标图像，将输入图像多层分解为基础层、纹理层和细节层。针对纹理层，采用基于对数Gabor滤波器的相位一致性（LGF-PC）方法计算显著性；针对细节层，采用常规Gabor滤波器计算显著性，并借助时域目标的显著性图来增强红外图像的显著性，从而实现对红外图像中气云目标的增强。最终，实现高质量的融合图像。实验表明：与传统融合算法相比，所提方法能够显著提升气云目标融合的增强效果，且在空间细节纹理保留能力上具备一定优势，能够在保留场景细节的同时提升气体图像质量。进一步结合气云图像合成彩色化方法，可增强气云目标的彩色化质量。所提方法可有效解决现有图像融合方法处理非显著目标时存在的缺陷，为基于图像融合技术的气体监测、弱小目标监测等领域提供了新的融合策略，具有广阔的应用前景。更多还原

【Abstract】 Objective Hazardous gas leak detection is of paramount importance in industrial safety, environmental protection, and emergency response. In practical applications, gas leak detection is typically conducted using visible and infrared dual-band imaging systems; however, the direct application of traditional fusion methods to these systems often leads to the degradation of key gas cloud target information. To address this problem, this paper proposes a dual-band gas image fusion enhancement method based on multi-scale decomposition and spatiotemporal saliency, named MSSTS. The method first extracts the gas cloud target in the temporal domain by analyzing infrared image sequences, and then utilizes multi-scale filtering to decompose the visible, infrared, and temporal target images into a base layer, a texture layer, and a detail layer. For the texture layer, saliency is computed using a phase congruency method based on Log-Gabor filters(LGF-PC), while for the detail layer, conventional Gabor filters are used. The saliency map of the infrared image is enhanced by leveraging the saliency map of the temporal target image, thereby enhancing the gas cloud target in the infrared image and ultimately realizing a high-quality fused image. Experiments show that, compared to traditional fusion algorithms, the MSSTS method achieves significant improvements in the fusion enhancement of the gas cloud target and demonstrates advantages in preserving spatial details and textures. It can enhance the quality of the gas image while preserving scene details. When combined with gas cloud synthesis and colorization methods, the quality of the colorized gas cloud target can be further enhanced. This paper resolves the deficiencies of existing image fusion methods in processing non-salient targets, providing a new fusion strategy for fields such as gas monitoring and weak target detection based on image fusion technology, with broad application prospects.Methods We propose a novel fusion enhancement method named MSSTS. Its core idea is to integrate temporal priors into a multiscale fusion architecture. The method proceeds in four main stages. First, we capture the target’s temporal dynamics. The robust SubSENSE algorithm is applied to the input infrared video sequence to isolate the moving gas plume, generating a temporal motion target image that is largely free from static background interference. Second, we perform a synchronized multi-scale decomposition. The visible frame, infrared frame, and the motion target image are each decomposed into a base layer(low-frequency structures), a texture layer(detailed patterns), and a detail layer(fine edges) using the edge-preserving rolling guidance filter(RGF). This allows for targeted processing of information at different scales. Third, we implement a temporally-guided saliency enhancement. A tailored dualsaliency strategy is adopted: a standard Gabor filter bank calculates saliency for the detail layers, while a more robust, contrastinvariant phase congruency(LGF-PC) method is used for the complex texture layers. The key innovation lies here: the saliency maps derived from the gas motion map are used to directly enhance the saliency maps of the infrared image. This is achieved through a fusion weighting scheme where temporal saliency acts as a modulator, ensuring that image regions exhibiting motion are given significantly higher priority during the fusion process. Finally, the layers are fused and reconstructed. The texture and detail layers are fused using a weighted-averaging strategy guided by the enhanced saliency maps. The base layers are simply averaged. The final enhanced image is obtained by summing the fused layers.Results and Discussions We conducted extensive experiments on a self-collected dual-band gas leak dataset and the public VIFB dataset. The qualitative results on gas leak scenes are compelling(Figs. 6-8). Unlike mainstream methods which visibly weaken the gas plume, MSSTS successfully enhances its contrast and structure against complex backgrounds. In challenging scenarios, such as distinguishing hot water vapor from the target gas, MSSTS effectively preserves the characteristics of both(Fig. 7). The proposed framework also enables a “Render-then-Fuse” strategy, allowing for a highly conspicuous, pseudo-colored gas visualization without the false alarms that plague traditional overlay methods(Fig. 11). Quantitative analysis(Table 1) corroborates the visual findings. On the gas dataset, MSSTS achieves a superior balance across various metrics. While some methods may score higher on a single metric at the cost of introducing artifacts, our approach demonstrates strong, well-rounded performance in information entropy(EN), edge information(EI), and visual quality(QCV), indicating effective target enhancement while preserving natural scene quality. Performance on the VIFB dataset further proves the method’s versatility. Computationally, MSSTS is efficient, with a GPUaccelerated processing time of 0.927 seconds per frame(640×512), demonstrating a clear path toward real-time application(Table 2).Conclusions This paper proposes an infrared image fusion enhancement method, MSSTS, based on image decomposition via multiscale guided filtering and spatiotemporal saliency. It acquires the gas cloud target from infrared images using a motion saliency detection method and employs multi-scale decomposition to synchronously decompose the infrared, visible, and temporal gas cloud target images. The multi-scale saliency of the temporal gas cloud target is then used to enhance the saliency of the infrared image, achieving an effective fusion of the visible and infrared images. When combined with a gas cloud synthesis and color rendering method, the quality of the colorized gas cloud target can be further enhanced, improving the visibility of the gas leakage area. The experimental results demonstrate that the MSSTS method achieves excellent fusion enhancement effects in the gas leakage image fusion task and is superior to other comparative methods in both subjective visual perception and objective evaluation metrics, effectively resolving the deficiencies of existing image fusion methods when applied to the fusion of non-salient targets. This research provides a new concept and method for image fusion in the multi-band imaging detection of gas leak clouds. In the future, through further optimization and improvement, the MSSTS method can be transplanted to a hardware platform for real-time operation, enabling it to play a greater role in practical applications such as gas leak detection and weak target detection.更多还原