Deep Guided Filtering and Detail Recovery-Based Infrared and Visible Image Fusion

To address the issues of poor detail clarity and limited visibility in infrared and visible image fusion under nighttime scenarios, this paper proposes a fusion method based on deep guided filtering and detail recovery. To fully exploit the multiscale features of source images, a multilevel image decomposition model is developed using deep guided filtering and Gaussian filtering, which further decomposes the high-frequency and low-frequency layers obtained by conventional multiscale decomposition into texture, saliency, and base information. In nighttime scenarios, these decomposed components typically suffer from blurring, low contrast, and insufficient brightness, making conventional fusion strategies inadequate for achieving satisfactory results. To cope with these challenges, three fusion strategies specifically tailored for nighttime imaging are devised: a texture information fusion strategy based on an enhancement function, a saliency information fusion strategy leveraging multi-contrast weights, and a base information fusion strategy that integrates the intensity transformation function of infrared images with grayscale difference analysis. To validate the performance of the proposed method, subjective and objective comparative evaluations are conducted on typical datasets including TNO, RoadScene, LLVIP, and MSRS, against classical fusion methods, state-of-the-art approaches, and methods specifically designed for nighttime imaging. Subjectively, the fused images produced by the proposed method exhibit richer and clearer details, along with more prominent thermal targets. Objectively, the proposed method demonstrates consistent advantages across eight metrics: average pixel intensity, standard deviation, average gradient, entropy, spatial frequency, total loss of information measure, perceptual saliency-based metric, and visual information fidelity. Both subjective and objective experimental results confirm that the proposed method can effectively accomplish nighttime image fusion tasks.