Semantic Prompt Enhancement for Semi-Supervised Low-Light Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing salient object detection (SOD) models struggle in low-light conditions due to inadequate training data and feature integration limitations.
• Low-light scenes present significant challenges for accurate data annotation, hindering the development of robust SOD models.

Purpose of the Study:

• To develop an effective semi-supervised framework for low-light salient object detection (SOD).
• To address limitations in current SOD models by enhancing contextual information and mitigating annotation burdens in low-light environments.

Main Methods:

• A brightness Retinex enhancer (BRE) was designed to mitigate illumination's impact on SOD tasks.
• A semi-supervised framework utilized sparse labeled semantic prompts to augment unlabeled data, incorporating Retinex decomposition and a context-guided encoder (CGE).
• Joint consistency training was performed between shared and perturbation decoders on both labeled and unlabeled data.

Main Results:

• The proposed semi-supervised model significantly enhances low-light SOD performance.
• The approach alleviates the substantial data annotation burden associated with low-light conditions.
• Experimental results show highly competitive performance compared to state-of-the-art fully supervised SOD models across multiple datasets.

Conclusions:

• The developed semi-supervised framework offers a promising solution for low-light salient object detection.
• This method effectively balances performance enhancement with reduced annotation requirements, making it practical for real-world low-light scenarios.