GANSD: A generative adversarial network based on saliency detection for infrared and visible image fusion

Abstract

Image fusion technology, which integrates infrared images providing valuable contrast information with visible light images rich in texture details, represents an effective and rational approach for object detection and tracking. Previous methods have often neglected crucial information due to a lack of saliency detection and have failed to fully utilize complementary information by separately processing different features from the two original images. To address these limitations and enhance fusion techniques, we propose a generative adversarial network with saliency detection (GANSD) for image fusion through an adversarial process. This approach simplifies the design of fusion rules and improves the quality of fused images. By incorporating saliency detection, GANSD effectively preserves both foreground and background information from the input images. The architecture also integrates complementary information to prevent data loss from the input images. Simultaneously, an attention mechanism within the generator emphasizes the importance of different feature channels. Extensive experiments on two public datasets, TNO and Roadscene, demonstrate that GANSD provides both qualitative and quantitative advantages over nine state-of-the-art (SOTA) methods.