Unsupervised saliency detection via multi-focus image reconstruction and prior-guided mask based on light field imaging

Recently, light field data has garnered significant attention due to its immense potential in Unsupervised Salient Object Detection (USOD). However, these methods neglect the ability of the light field information itself to generate pseudo-labels. In this paper, we design a two-stage pseudo-label generation framework, based on the data structure of light field. In the first stage, we propose a proxy task called multi-focus image reconstruction (MFIR). It leverages light field information to generate a shallow depth-of-field image with the focus on the salient object, approximating the learning of saliency features. In the second stage, we introduce repair network and prior-guided mask (PGM) to guide pseudo-label updating by leveraging the stability of salient features in pre-trained weights, thereby addressing the depth ambiguity issue arising from MFIR. We name our framework light field refocus for saliency (LFR4S). Additionally, we use the generated pseudo-labels for supervised training and conduct comparative analysis on the results. Experimental results demonstrate that our method surpasses most existing USOD methods across multiple datasets. Finally, we design corresponding ablation studies to verify the necessity of certain modules. • Propose a proxy task to extract light field saliency features without supervision. • Propose a mask update strategy to optimize performance of pre-trained network. • Design a module for better fusing focal stack and depth features.