Spectrum-oriented Point-supervised Saliency Detector for Hyperspectral Images

Abstract

Hyperspectral salient object detection (HSOD) aims to extract targets or regions with significantly different spectra from hyperspectral images. While existing deep learning-based methods can achieve good detection results, they generally necessitate pixel-level annotations, which are notably challenging to acquire for hyperspectral images. To address this issue, we introduce point supervision into HSOD, and incorporate Spectral Saliency, derived from conventional HSOD methods, as a pivotal spectral representation within the framework. This integration leads to the development of a novel Spectrum-oriented Point-supervised Saliency Detector (SPSD). Specifically, we propose a novel pipeline, specifically designed for HSIs, to generate pseudo-labels, effectively mitigating the performance decline associated with point supervision strategy. Additionally, Spectral Saliency is employed to counteract information loss during model supervision and saliency refinement, thereby maintaining the structural integrity and edge accuracy of the detected objects. Furthermore, we introduce a Spectrum-transformed Spatial Gate to focus more precisely on salient regions while reducing feature redundancy. We have carried out comprehensive experiments on both HSOD-BIT and HS-SOD datasets to validate the efficacy of our proposed method, using mean absolute error (MAE), E-measure, F-measure, Area Under Curve, and Cross Correlation as evaluation metrics. For instance, on the HSOD-BIT dataset, our SPSD achieves a MAE of 0.031 and an F-measure of 0.878. Thorough ablation studies have substantiated the effectiveness of each individual module and provided insights into the model's working mechanism. Further evaluations on RGB-thermal salient object detection datasets highlight the versatility of our approach.