Seeing the Intangible: Survey of Image Classification into High-Level and Abstract Categories

A comprehensive survey paper examines how computer vision systems classify images into high-level and abstract categories, revealing that current approaches struggle with conceptual understanding beyond simple visual features. The research identifies key challenges including dataset limitations and the need for hybrid AI systems that integrate supplementary information to better handle abstract concepts like emotions, aesthetics, and ideologies.

This survey addresses a fundamental gap in computer vision research by systematizing approaches to high-level visual understanding—a shift from traditional object detection toward more nuanced semantic interpretation. The paper categorizes abstract concept (AC) classification into distinct semantic clusters including commonsense reasoning, emotional perception, aesthetic judgment, and interpretative analysis. This taxonomy matters because it clarifies what researchers mean by 'understanding' when systems analyze images at conceptual rather than pixel levels.

The research reveals that the field has reached an inflection point where scale alone—massive datasets and larger models—no longer suffices for abstract reasoning tasks. This finding contradicts the prevailing deep learning narrative that bigger models solve harder problems. Instead, the survey emphasizes that successful AC classification requires hybrid approaches combining multiple AI methodologies with external knowledge sources and mid-level feature representations that capture meaningful patterns between raw pixels and high-level concepts.

For the AI development community, this survey implies significant methodological shifts ahead. Teams building vision systems for content moderation, sentiment analysis, or cultural interpretation cannot rely solely on supervised learning from labeled images. The integration of symbolic reasoning, knowledge graphs, and human-in-the-loop validation becomes critical. This complexity increases development costs and timelines for AI products targeting abstract reasoning, potentially creating competitive advantages for teams willing to invest in sophisticated engineering beyond standard deep learning architectures. The research establishes academic groundwork that will influence production system design across industries relying on nuanced image understanding.

• →High-level image classification requires hybrid AI systems combining multiple approaches, not just scaling existing deep learning models.
• →Current large datasets have diminishing returns for abstract concept classification tasks requiring emotional, aesthetic, or ideological understanding.
• →Computer vision research must integrate supplementary information and mid-level features to move beyond surface-level visual pattern recognition.
• →Abstract concept classification spans commonsense reasoning, emotional perception, aesthetics, and interpretation—requiring different technical approaches.
• →The field is shifting from object detection toward conceptual visual sensemaking, fundamentally changing how vision systems should be architected.