Select Page

Dataset Description

Conclusions and Recommendations

Summary of Key Findings

This analysis provides a comprehensive assessment of visual diversity, semantic content, and representational bias across a dataset of 7,362 images spanning 20 conceptual categories.

Key findings include:

  • A balanced and semantically rich dataset, capturing over 113,000 object instances across psychological, developmental, and cultural themes.
  • Statistical correction of representational bias through inverse category size weighting, achieving near-perfect analytic equality without altering the raw data.
  • Strong empirical support for Shannon Diversity Index as a valid and robust measure of visual complexity, with abstract categories consistently exhibiting higher diversity.
  • A predictive framework in which unique object count and object richness ratio explain over 91% of variation in visual diversity.

Methodological Contributions

This project demonstrates a high standard of analytical design for image-based research:

  • Empirical validation of Shannon Diversity Index through alternative metrics and correlation convergence.
  • Rigorous assumption testing (normality, homoscedasticity, autocorrelation) and strong model fit (R² = 0.917).
  • Application of multiple diversity metrics and thorough bias correction procedures tailored for computer vision outputs.

Limitations 

  • Category Scope: Only 20 categories were included, limiting statistical power for smaller or emerging visual patterns.
  • Detection Constraints: Automated object detection, while high-performing, may overlook spatial and aesthetic nuances not captured in object frequency.
  • Temporal & Cultural Gaps: Static image snapshots do not account for time-dependent or culturally contingent shifts in visual representation.

Recommendation

To ensure fair and interpretable comparisons, inverse category size weighting and object normalisation should be retained throughout analysis. Shannon Diversity Index remains the most reliable measure of visual complexity, particularly for distinguishing abstract from concrete categories.

Category comparisons should be treated cautiously, especially where content is visually stereotyped (e.g., education, nature). All findings are intended to be valid within the dataset as constructed, not generalised beyond it.

Maintaining a modular workflow is recommended to support future analysis or adaptations without disrupting core methods.

←→