4DLatLRR Combined with Dual-Path Generative Adversarial Network for Infrared and Visible Image Fusion

The use of deep learning techniques for infrared and visible image fusion has produced numerous outstanding research results. However, challenges remain in preserving the source image information, as issues such as imbalances and insufficient feature extraction can lead to information distortion or visually unnatural results. In this study, a method combining a four-layer deep latent low-rank representation with a dual-path generative adversarial network for image fusion is proposed. First, the source image is decomposed into low-rank and salient images. The low-rank layer uses a dual-path generative adversarial network with multiple classifications to estimate the distributions of the infrared and visible low-rank regions simultaneously. The salient layer employs a fusion strategy based on the nuclear norm to flexibly retain and highlight local textures, while introducing gradient images to enhance the detailed information. In addition, primary and auxiliary loss functions are designed to constrain the generator during the extraction of gradient and intensity information. During intensity extraction, the infrared image is primarily used to complement the limited intensity information in the visible image, whereas during gradient extraction, the visible image is primarily used to supplement the limited gradient information in the infrared image. The experimental results demonstrate that this method significantly outperforms the others across five key metrics, with the overall difference correlation sum indicator ranking at the upper-middle level.