Weighted multi-scale and wavelet-enhanced Segment Anything Model for salient object detection

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Salient Object Detection (SOD) aims to identify and segment visually prominent objects.
• Existing Segment Anything Model (SAM) adaptations for SOD struggle with incomplete masks, blurred boundaries, and fine details, especially in complex scenes.
• Limitations include rigid multi-scale fusion, ignored cross-level semantic correlations, and loss of high-frequency details in spatial-domain operations.

Purpose of the Study:

• To address the limitations of current SOD methods, particularly those based on SAM.
• To propose a novel model, WMW-SAM, for accurate and detailed salient object detection.
• To improve segmentation of objects with diverse scales and complex textures.

Main Methods:

• Developed a Weighted Multi-Scale Adapter (WMSA) for learnable cross-scale feature calibration.
• Introduced a Multi-level Feature Cross-fusion Module (MFCM) using cross-attention for semantic-detail interaction.
• Implemented a Detail Enhancement Module (DEM) with Discrete Wavelet Transform (DWT) to recover high-frequency details and sharp boundaries.

Main Results:

• WMW-SAM demonstrated superior performance on multiple benchmark datasets for salient object detection.
• The model achieved accurate saliency predictions with enhanced fine-grained details.
• WMW-SAM effectively recovered sharp object boundaries and intricate textures often missed by other methods.

Conclusions:

• WMW-SAM effectively overcomes the limitations of existing SAM adaptation strategies for SOD.
• The proposed approach significantly improves the quality and detail of saliency maps.
• WMW-SAM offers a robust solution for accurate and detailed salient object detection in challenging visual scenes.