Dynamic fusion method for infrared-visible images based on dual-guided filtering and a parallel CNN

Infrared and visible image fusion technology is widely applied in military reconnaissance, security surveillance, and power equipment inspection. However, traditional methods rely on manual feature extraction, struggling to adaptively separate low-frequency thermal radiation and high-frequency texture information in multimodal images. Deep learning approaches often neglect edge consistency, leading to blurred thermal boundaries and detail loss in fused images. To address these issues, this paper proposes a fusion framework based on the collaborative optimization of dual-guided filtering (Dual-GF) and a parallel convolutional neural network (PCNN). Dual-GF independently extracts low-frequency base layers and high-frequency detail layers from infrared and visible images to avoid modal confusion. A lightweight PCNN architecture is designed: shared parameters at lower layers extract common features, while higher layers employ dilated convolutions in the infrared branch and skip connections in the visible branch to capture thermal diffusion patterns and texture details. A channel-spatial attention mechanism (CSAM) is proposed to achieve dynamic fusion, which adaptively assigns pixel-level weights to infrared thermal targets and visible textures based on their saliency. An improved multi-scale gradient consistency loss (MGCL) is proposed to jointly optimize gradient magnitude and directional consistency, suppressing structural distortion. The method is compared with six other algorithms, including DenseFuse and FusionGAN, using the TNO dataset. The experimental results show that the proposed method improves PSNR, VIF, EN, and SF by an average of 3.12%, 28.46%, 6.25%, and 31.96%, respectively, while achieving the lowest MSE and SCD, demonstrating its superiority in preserving thermal target saliency and texture details.