RGB-D salient object detection via convolutional capsule network based on feature extraction and integration.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Fully convolutional neural networks (FCNNs) excel at salient object detection using RGB or RGB-D data but struggle with object-part segmentation.
• Capsule networks can identify complete objects but are computationally intensive and time-consuming.
• A need exists for efficient methods that can accurately segment salient objects while preserving object-part relationships.

Purpose of the Study:

• To propose a novel convolutional capsule network (CNN) for salient object detection that resolves the object-part dilemma.
• To develop a method with reduced computational demand compared to traditional capsule networks.
• To improve the accuracy and completeness of salient object segmentation.

Main Methods:

• Utilized a VGG backbone for RGB feature extraction and integration.
• Incorporated a feature depth module to fuse RGB features with depth image information.
• Employed a feature-integrated convolutional capsule network with locally-connected routing for object-part relationship exploration.
• Generated the final salient map using deconvolutional capsules.

Main Results:

• The proposed method effectively addresses the object-part dilemma in salient object detection.
• Achieved superior performance compared to 23 state-of-the-art algorithms on four RGB-D benchmark datasets.
• Demonstrated reduced computational demand while maintaining high accuracy.

Conclusions:

• The novel convolutional capsule network offers an effective solution for salient object detection, overcoming limitations of FCNNs and traditional capsule networks.
• The method demonstrates significant improvements in segmentation completeness and accuracy.
• This approach presents a computationally efficient and high-performing alternative for salient object detection tasks.