A neuro-fuzzy approach for segmentation of human objects in image sequences

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Object segmentation is crucial for high compression ratios in video coding standards like MPEG-4 and MPEG-7.
• Human objects are primary components in multimedia video streams, necessitating accurate segmentation.
• Current segmentation methods often suffer from restricted usage or high error rates due to simplistic detection criteria.

Purpose of the Study:

• To develop a novel and accurate approach for segmenting human objects (face and body) in image sequences.
• To overcome limitations of existing methods by integrating temporal and spatial information with a neuro-fuzzy mechanism.

Main Methods:

• Utilizing fuzzy self-clustering to segment the base frame into foreground and background categories based on multiple criteria.
• Employing a fuzzy neural network, trained via a singular value decomposition (SVD)-based hybrid learning algorithm, for precise human object localization.
• Combining temporal and spatial information for enhanced segmentation accuracy.

Main Results:

• The proposed neuro-fuzzy approach demonstrates superior segmentation performance compared to existing methods.
• Accurate localization of human objects in both base and subsequent frames of video streams.
• Effective categorization of image segments into foreground and background.

Conclusions:

• The novel neuro-fuzzy approach offers a significant improvement in human object segmentation for video sequences.
• This method enhances the efficiency and accuracy of video compression and multimedia applications.
• The integration of fuzzy logic and neural networks provides a robust solution for complex segmentation tasks.