Diving in a bit. There’s a lot of different ways to measure aesthetics. It’s interesting to see all the different algorithmic approaches.
Here’s a quick overview from Perplexity.
Answer from Perplexity:
Ways to Measure Aesthetics in Computational and Experimental Research
This comprehensive report outlines the various methods and approaches used to measure aesthetics as described in the research article from Frontiers in Computational Neuroscience. The intersection of computational aesthetics and experimental aesthetics has produced diverse techniques for quantifying what makes images aesthetically appealing.
Historical Approaches to Aesthetic Measurement
Early Mathematical Formulations
The foundations of computational aesthetics can be traced back to Birkhoff's 1933 book "Aesthetic Measure," where he proposed a mathematical formula for aesthetic value. Birkhoff's formula expressed aesthetic measure (M) as a function of order (O) and complexity (C), specifically M = O/C, suggesting that aesthetic reward should be proportional to effort. This early approach attempted to quantify aesthetics through a simple mathematical relationship between these two fundamental properties.
Fechner's Empirical Aesthetics
Gustav Fechner, the founder of experimental aesthetics, established a scientific approach to measuring aesthetic appeal through stimulus properties in his 1876 book "Vorschule der Ästhetik". Fechner believed aesthetic appeal could be measured objectively through formalistic properties of physical objects, such as the golden ratio in rectangles. His approach, called "aesthetics from below," provided the foundation for empirical studies that relate physical image properties to aesthetic perception.
Feature-Based Aesthetic Measurement
Handcrafted Features
Early computational aesthetic systems relied on purposely designed features that captured specific properties thought to influence aesthetic appeal:
Compositional features: Measurements based on photographic principles like the rule of thirds, which quantifies how well an image follows established compositional guidelines.
Color features: Various color measurements including colorfulness, saturation, hue distribution, and color harmony. Researchers like Datta et al. (2006) found that average hue and a saturation-based rule of thirds measure contributed significantly to aesthetic value.
Clarity measures: Features capturing image sharpness, blur, and contrast. Ke et al. (2006) found blur to be the most discriminative metric for distinguishing high-quality from low-quality photographs.
Brightness and lighting: Measures of average pixel intensity, lighting distribution, and brightness patterns.
Subject-background distinction: Luo and Tang (2008) improved classification by extracting features separately from subject regions and backgrounds rather than from whole images.
Generic Image Features
As computational aesthetics evolved, researchers began using more general-purpose features that weren't specifically designed for aesthetic evaluation:
SIFT descriptors: Scale-invariant feature transform features originally designed for feature matching but repurposed for aesthetic evaluation.
Color descriptors: Generic representations of color information that capture properties like color saturation without being explicitly designed for aesthetics.
Bag-Of-Visual-Words: A method that represents images as collections of visual "words" to capture prototypical patches for aesthetic and non-aesthetic photographs.
Fisher-Vector representation: A statistical framework for encoding images, which Marchesotti et al. (2011) used to improve classification rates over handcrafted features.
Advanced Computational Approaches
Deep Learning Methods
Recent advances in computational aesthetics have been dominated by deep learning approaches:
Convolutional Neural Networks (CNNs): These networks learn hierarchical filters that extract meaningful information from images, capturing both local properties like sharpness and global properties like composition.
Multi-column architectures: Lu et al. (2015) developed a two-column deep neural network that simultaneously processes global and local views of photographs to capture both compositional and detail-specific aesthetic properties.
Fine-tuned networks: Networks originally trained for object recognition can be repurposed for aesthetic evaluation through fine-tuning, demonstrating that content recognition contributes to aesthetic judgment.
Multi-task learning: Kao et al. (2016) showed that networks trained to recognize both semantic content and aesthetic quality perform better than those trained solely for aesthetic classification.
Statistical Image Properties
Luminance and Contrast Properties
Several statistical properties related to brightness and contrast have been associated with aesthetic appeal:
Luminance skewness: Graham and Field (2007, 2008) found that artworks tend to have lower-skew luminance histograms than photographs of real scenes, with roughly equal proportions of light and dark areas.
Luminance fluctuations: Aragón et al. (2008) discovered that the distribution of luminance in some Van Gogh paintings resembled mathematical distributions of fluid turbulence, suggesting specific patterns may evoke particular emotional responses.
Edge characteristics: Leykin and Cutzu (2003) found that paintings contain significantly more color-only edges than photographs of real scenes, and that color edges and intensity edges coincide less frequently in paintings.
Complexity Measurements
Complexity has been extensively studied as a factor in aesthetic appreciation:
Element counting: Simply counting the number of visual elements in an image, as proposed by Birkhoff (1933).
Fractal dimension: A measure of self-similarity across scales, used to quantify the complexity of both natural scenes and artworks.
Compression-based measures: GIF compression has been used as a proxy for visual complexity by Forsythe et al. (2011).
Edge density: The overall amount of edges in an image has been used to quantify visual complexity by Redies et al. (2017).
Information theory measures: Rigau et al. (2008) proposed measures based on information theory and Kolmogorov complexity to quantify order versus complexity in paintings.
Symmetry Analysis
Symmetry plays a significant role in aesthetic perception, though its importance varies by domain:
Geometric symmetry detection: Measurements of reflectional, translational, or rotational symmetry in images.
Facial symmetry: Particularly relevant for evaluating attractiveness of human faces.
Compositional balance: While perfect symmetry is rare in artworks, measures of overall visual balance contribute to aesthetic judgments.
Domain-Specific Aesthetic Measurements
Artist and Style Identification Methods
Techniques developed to identify artists or artistic styles often capture aesthetic properties:
Discrete Cosine Transform coefficients: Used by Keren (2002) to identify painters with 86% accuracy.
Wavelet decomposition: Applied by Lyu et al. (2004) to distinguish authentic paintings from forgeries and identify multiple artists' contributions to a single work.
Brushstroke analysis: Methods that capture the distinctive patterns of an artist's brushwork.
Classeme features: Vectors representing combined classification scores from multiple weak classifiers trained on low-level descriptors, which proved effective for style classification.
Forgery Detection Techniques
Methods developed for art authentication often employ highly sensitive measurements:
Wavelet statistics: Lyu et al. (2004) extracted feature vectors from wavelet decompositions to distinguish authentic paintings from imitations.
Hidden Markov Models: Polatkan et al. (2009) used parameters from Hidden Markov Models trained on wavelet coefficients to discriminate copies from originals.
Fractal analysis: Taylor et al. (1999) performed fractal analysis of Jackson Pollock's drip paintings and found that the fractal dimension increased over the artist's lifetime, potentially providing a method for authentication.
Integrative Approaches
Multimodal Analysis
Some researchers combine multiple measurement approaches:
Combined visual and spectral analysis: Montagner et al. (2016) integrated brushstroke analysis using SIFT features with pigment analysis using hyperspectral imaging for forgery detection.
Content and form integration: Recent approaches recognize that both perceptual (form-based) and cognitive (content-based) factors contribute to aesthetic appreciation.
Cross-domain feature fusion: Combinations of low-level features (capturing form) with high-level semantic features (capturing content) often yield the best performance in aesthetic classification.
Conclusion
The measurement of aesthetics has evolved from simple mathematical formulations to sophisticated computational approaches that integrate multiple levels of visual analysis. While early methods focused on handcrafted features based on artistic intuition, recent advances leverage deep learning to capture both formal properties and semantic content that contribute to aesthetic appeal.
The diversity of measurement approaches reflects the complex, multifaceted nature of aesthetic experience, which involves perceptual, cognitive, and emotional components. Future measurements of aesthetics will likely continue to integrate these different aspects while developing more interpretable models that can explain, not just predict, what makes an image aesthetically pleasing.
Citations:
Answer from Perplexity