Boosting Broader Receptive Fields for Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient Object Detection (SOD) excels at regular-scale targets but struggles with scale variation.
• Existing SOD methods face performance limitations with extremely large or small objects due to insufficient receptive fields.

Purpose of the Study:

• To propose a novel framework, BBRF (Boosting Broader Receptive Fields), to address the challenge of scale variation in salient object detection.
• To enhance the ability of SOD models to perceive and segment objects across a wide range of scales, including extreme sizes.

Main Methods:

• Developed a Bilateral Extreme Stripping (BES) encoder to separate semantics and details, broadening receptive fields.
• Introduced a Dynamic Complementary Attention Module (DCAM) for dynamic filtering of bilateral features.
• Proposed a Switch-Path Decoder (SPD) guided by a Loop Compensation Strategy (LCS) and a boosting loss for feature enhancement.

Main Results:

• The BBRF framework demonstrates a significant advantage in handling scale variation in salient object detection.
• Achieved over 20% reduction in Mean Absolute Error (MAE) compared to state-of-the-art methods on five benchmark datasets.
• Successfully improved the perception and segmentation of extremely large and small objects.

Conclusions:

• BBRF effectively overcomes the limitations of existing SOD methods in processing scale-variant objects.
• The proposed BES encoder, DCAM, and SPD with LCS provide a robust solution for comprehensive receptive field acquisition.
• BBRF offers a promising advancement for salient object detection tasks requiring accurate segmentation across diverse object scales.