Visual complexity modelling based on image features fusion of multiple kernels

Area of Science:

• Computational aesthetics
• Machine learning
• Human-computer interaction

Background:

• Human perception of visual complexity is crucial for aesthetic order.
• Accurate computational models are challenging due to cognitive and neurological factors.
• Previous research has explored various computational approaches.

Purpose of the Study:

• To develop accurate computational models of visual complexity using machine learning.
• To compare different machine learning techniques for modeling visual complexity.
• To identify key features influencing human perception of visual complexity.

Main Methods:

• Utilized a dataset of 800 visual stimuli with 329 features (edge detection, compression error, Zipf's law).
• Performed comparative analysis of state-of-the-art machine learning algorithms.
• Conducted outlier analysis and evaluated the impact of outlier removal.

Main Results:

• Feature Selection Multiple Kernel Learning (FS-MKL) demonstrated superior performance.
• Achieved an average correlation of 0.71 with human perception of complexity.
• Identified a reduced set of 22 features that significantly improve model accuracy.

Conclusions:

• FS-MKL effectively models human perception of visual complexity.
• Outlier analysis and feature selection enhance model performance.
• This approach shows significant potential for regression tasks in computational aesthetics.