Depthwise Nonlocal Module for Fast Salient Object Detection Using a Single Thread.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks excel in salient object detection but demand high-end GPUs for real-time performance.
• Existing mobile-optimized architectures are unsuitable for salient object detection's unique correlation modeling needs.

Purpose of the Study:

• To develop a deep-learning algorithm for fast salient object detection.
• To achieve both competitive accuracy and high inference efficiency on a single CPU thread.

Main Methods:

• Proposed a novel depthwise nonlocal module (DNL) to model contrast by harvesting intrachannel and interchannel correlations using self-attention.
• Introduced a depthwise nonlocal network architecture integrating DNL modules and inverted residual blocks.

Main Results:

• The proposed algorithm achieves competitive accuracy on various salient object detection datasets.
• Demonstrated state-of-the-art inference efficiency among deep-learning-based algorithms on a single CPU thread.

Conclusions:

• The developed algorithm enables fast salient object detection on resource-constrained devices.
• The novel DNL module and network architecture effectively capture essential correlations for accurate and efficient salient object detection.