Image Visual Realism: From Human Perception to Machine Computation.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Image Processing

Background:

• Assessing visual realism is crucial for computer graphics and photo editing.
• Current methods rely on time-consuming human evaluation or reference-based algorithms.
• A need exists for efficient, reference-free visual realism prediction.

Purpose of the Study:

• To develop a reference-free computational framework for predicting visual realism.
• To identify image attributes that significantly influence human perception of realism.
• To create interpretable models of human visual realism perception.

Main Methods:

• Constructed a benchmark dataset of 2,520 images with human-annotated attributes.
• Performed statistical modeling to identify key attributes for visual realism.
• Developed empirical and deep convolutional neural network (CNN) models for realism prediction.

Main Results:

• Identified specific image attributes that are most relevant to visual realism.
• Developed an interpretable empirical model characterizing human realism perception.
• Created CNN models capable of predicting visual realism without reference images.

Conclusions:

• The proposed framework offers an efficient, reference-free approach to visual realism prediction.
• The study links computational features to latent factors in human image perception.
• This work advances automated assessment of image realism for various applications.