PoolNet

PoolNet improves salient object detection by integrating pooling-based Global Guidance and Feature Aggregation modules into a U-shape convolutional neural network to preserve spatial details and refine semantic features.

Key Features:

• Global Guidance Module (GGM): A Global Guidance Module based on the bottom-up pathway of a U-shape architecture directs location information of potential salient objects across feature levels to preserve spatial details.
• Feature Aggregation Module (FAM): A Feature Aggregation Module in the top-down pathway progressively fuses coarse-level semantic information with fine-level features to refine saliency maps.
• Performance and Efficiency: Accurately locates salient objects with enhanced detail and processes 300 × 400 images at 53 frames per second (FPS) compared to existing state-of-the-art methods.
• Mobile Optimization: Uses MobileNetV2 as the backbone and optimizes pooling-based modules for mobile deployment, achieving 66 FPS on mobile devices while maintaining performance that surpasses most existing methods.

Scientific Applications:

• Edge Detection: Identifies image edges to enhance boundary delineation and support image segmentation.
• RGB-D Salient Object Detection: Integrates RGB and depth (RGB-D) information to detect salient objects in three-dimensional space.
• Camouflaged Object Detection: Detects camouflaged objects from backgrounds by extracting and integrating detailed features.

Methodology:

Applies pooling techniques within a U-shape CNN using bottom-up and top-down pathways; the Global Guidance Module transfers location information across feature levels and the Feature Aggregation Module fuses coarse-level semantic information with fine-level features; MobileNetV2 is used as a backbone for mobile optimization.