CsdlFusion: An Infrared and Visible Image Fusion Method Based on LatLRR-NSST and Compensated Saliency Detection

Abstract

Image fusion methods may lose their ability to retain crucial image information when faced with suboptimal conditions, such as poor contrast, excessive noise, or intense illumination, leading to the loss of valuable image features. In this work, an improved CsdFusion algorithm is proposed to increase the visibility of infrared targets in fused images. Firstly, to accomplish clear background textures and structural information, a hybrid image decomposition model combining LatLRR and NSST is established. This process entails the division of the original infrared and visible images into low-rank components (base layers) and salient components (saliency layers) through the Latent Low-Rank Representation (LatLRR) approach. Subsequently, the base layers of both the infrared and visible images undergo the Non-Subsampled Shearlet Transform (NSST), decomposing them into high-frequency and low-frequency layers. The processed high-frequency and low-frequency layers are then subjected to inverse NSST to obtain the fused base layer, ensuring that the fused image retains maximum background information while effectively filtering noise. Secondly, to identify and extract the most significant regions or features in infrared images, the Central-contrast priori Saliency Map (CSM) algorithm is applied. This algorithm calculates the central prior saliency value using Harris corners and the contrast prior saliency value using guided filtering and background suppression. It then combines these two prior saliency values using a feature compensation strategy to compute the infrared saliency map. To validate the effectiveness of the proposed algorithm, comparative evaluation studies on benchmark open datasets are carried out. The results thus obtained through the proposed algorithm demonstrate superior performance in both subjective and objective experiments, generating fused images that not only preserve the crucial details and characteristics of both infrared and visible images but also reflect significant enhancement in visibility and discriminability of target objects, outperforming 10 state-of-the-art image fusion algorithms.