Low Level Texture Features for Snore Sound Discrimination.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Snoring is linked to serious health issues like obstructive sleep apnea and heart disease.
• Accurate snore sound analysis is crucial for identifying underlying causes and potential interventions.
• Automated, non-invasive methods for analyzing snore sound origins are increasingly important.

Purpose of the Study:

• To investigate the effectiveness of low-level image texture features for classifying four specific types of snore sounds.
• To compare the performance of Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG) in snore sound characterization.
• To evaluate the combined performance of LBP and HOG features and compare them against existing methods.

Main Methods:

• Utilized color spectrograms of snore sounds as input data.
• Extracted texture features using Histogram of Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG).
• Employed Support Vector Machines (SVM) with homogeneous mapping for snore sound classification.

Main Results:

• Local Binary Patterns (LBP) outperformed Histogram of Oriented Gradients (HOG) in classifying snore types.
• Fusion of LBP and HOG features yielded superior results compared to individual descriptors.
• The proposed method demonstrated relative percentage increases in unweighted average recall of 23.1% and 8.3% over the challenge baseline and deep spectrum features, respectively.

Conclusions:

• Low-level image texture features, particularly LBP, are effective for snore sound classification.
• Combining LBP and HOG features enhances classification performance for snore sounds.
• This approach offers a promising non-invasive method for analyzing snore sound characteristics and identifying potential health risks.