RGB-T Salient Object Detection Based on the Segment Anything Model
Shuohao He, Lin Shi · 2024
Salient object detection mimics the human visual system, segmenting targets using RGB and thermal infrared images suitable for low-light and occlusion scenarios. This technology faces challenges such as mismatched image resolutions and high equipment costs. This paper introduces an RGB-T salient object detection algorithm based on the “Segment Anything Model” (SAM), which enhances detection performance through modules like “High-Resolution Transformer Pixel Extraction” (HRTP), “Multiple Amplification Fusion” (MAF), “Noise Reduction Fusion Module” (NRF), and “Entropy-Based Pixel-Level Weighting Module” (EBPW). Experimental results show that this algorithm performs excellently on major datasets, accurately detecting salient objects and demonstrating potential for practical applications.