AI 생성 이미지의 진위 판별을 위한 시각적 패턴 분석 연구

초록

Conventional AI-generated image detection methods focus on technical accuracy, but often fail to incorporate cognitive factors such as visual discomfort perceived by humans. To address this limitation, this study proposes an authenticity detection framework based on a perceptual evaluation that reflects human visual cognition, emphasizing that such an approach should complement existing technical detection methods. Existing detection techniques primarily rely on pixel-level analysis, frequency transformation, and neural network-based models. These methods assess authenticity by learning the statistical signals and artifacts of generative algorithms. However, they are highly dependent on training data and show significantly reduced performance when faced with newly developed algorithms, especially diffusion-based models. In response, this study analyzes perceptual inconsistencies in AI-generated images by focusing on three key visual elements: lighting, image composition, and perspective. The results reveal a range of visual artifacts, including mismatches in lighting direction and color temperature, physically implausible interactions among multiple light sources, imbalance in surface texture, unnatural object overlap, and irregular pixel density. Perspective-related errors such as distorted object proportions, unnatural depth, and broken symmetry in visual flow were also identified. Based on this analysis, the study proposes a perceptual cue-based framework that reflects human visual intuition and compensates for the limitations of purely technical detection methods. Future research should advance toward a more systematic exploration that integrates the diversity of generative models with emerging patterns of visual inconsistencies.

키워드