NRA-Net-Neg-Region Attention Network for Salient Object Detection with Gaze Tracking

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Existing deep learning methods for salient object detection often suffer from feature loss due to autoencoder structures.
• This feature loss can lead to inaccurate detection results, either missing objects or identifying incorrect areas.

Purpose of the Study:

• To propose a novel detection method for salient objects, specifically for gaze tracking applications, that overcomes the limitations of existing autoencoder-based approaches.
• To develop a deep learning model that minimizes feature loss and enhances the accuracy of object detection in single images.

Main Methods:

• A network utilizing Neg-Region Attention (NRA) was developed to predict objects with a concentrated line of sight.
• The NRA method separates positive and negative regions using an exponential linear unit activation function, followed by region-specific attention.
• This attention mechanism, independent of a backbone network, emphasizes object areas while suppressing background noise.

Main Results:

• The proposed NRA method demonstrated reduced feature loss compared to conventional deep learning techniques.
• Experimental results showed that NRA achieved higher detection accuracy for salient objects.
• The method effectively emphasized target object areas and suppressed background regions.

Conclusions:

• The Neg-Region Attention (NRA) method offers a significant improvement in salient object detection for gaze tracking.
• NRA provides a more accurate and robust solution by mitigating feature loss inherent in traditional autoencoder models.
• This approach enhances the precision of identifying objects of interest in single images without requiring additional hardware.