Salient Object Detection With Purificatory Mechanism and Structural Similarity Loss.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning models excel at salient object detection using attention mechanisms.
• Current methods struggle with small, complex, or indistinguishable regions, leading to mispredictions.
• Accurate salient object detection is crucial for various image analysis tasks.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) for improved salient object detection.
• To address limitations in detecting small, complex, and indistinguishable salient regions.
• To enhance the accuracy and detail of salient object predictions.

Main Methods:

• Proposed a novel CNN incorporating a 'purificatory mechanism' with 'promotion attention' and 'rectification attention'.
• Promotion attention focuses on salient regions, while rectification attention corrects errors in difficult areas.
• Introduced 'Structural Similarity Loss' to model region-level relationships and refine saliency values in complex structures.

Main Results:

• The purificatory mechanism and Structural Similarity Loss significantly improved detection performance.
• The proposed method outperformed 19 state-of-the-art approaches across six benchmark datasets.
• Achieved efficient processing speeds of over 27 FPS on a single NVIDIA 1080Ti GPU.

Conclusions:

• The novel CNN with purificatory mechanism and Structural Similarity Loss effectively enhances salient object detection.
• The approach accurately predicts locations and details, particularly in challenging regions.
• Demonstrated superior performance and efficiency compared to existing state-of-the-art methods.