| In recent years,image fusion technology has received increasing attention and research,especially in the field of infrared and visible image fusion.Infrared images have advantages in low light,nighttime,and foggy environments,while visible images can provide more detail and color information.Combining these two types of images can produce images with higher quality and more comprehensive information,which has important applications in military,security,medical,environmental monitoring,and other fields.Traditional fusion methods rely on manually designed features and models,and their effectiveness and performance are limited.In recent years,infrared and visible image fusion based on deep learning technology has become a research hotspot.The powerful feature extraction and learning capabilities of deep learning models can achieve higher performance and effectiveness.However,there are still problems such as unstable fusion image quality,imbalanced information,and slow fusion speed.To solve these problems,two fusion methods based on generative adversarial networks were designed in this paper,namely Dual-frequency Cross-enhanced Fusion network(DCFusion)and Equi-frequency Division Convolution Real-time Fusion Network(EFD-Fusion).DCFusion utilizes a frequency decomposition and enhancement module based on the Laplacian of Gaussian to design a dual-frequency cross-enhancement fusion generator network,which achieves enhanced fusion.Compared with most existing advanced fusion algorithms,this method performs better in qualitative and quantitative comparison,as well as object detection accuracy.Even in extreme scenarios such as complex lighting,low light,and smoke,DCFusion can effectively enhance the target and background information of the scene to achieve enhanced fusion.EFD-Fusion designs a lightweight equal-scale frequency division convolution(EFDConv)based on Oct Conv to achieve real-time fusion.Experimental results show that the fusion results of EFD-Fusion can not only preserve the texture details of visible images but also retain the prominent targets in infrared images.Moreover,while ensuring fusion quality,this model greatly saves computational resources and improves processing speed.Compared with ten other advanced image fusion methods,the fusion speed of this method increases by an average of 89.7% on the TNO dataset,99.4% on the Road dataset,and 97% on the MSRS dataset.Both methods are end-to-end models,avoiding the need for complex manual fusion rules design. |
