서울 주요 카페 지역의 감성적 특성과 고객 평가의 상관관계 분석 -멀티모달 딥러닝 기반 감성 이미지 분석과 정량적 검증-

Abstract

(Background and Purpose) Commercial spaces increasingly function as emotional and experiential platforms that shape brand identity and consumer behavior beyond their traditional roles as sites of consumption. In particular, café environments leverage visual stimuli to influence customers’ initial impressions, emotional responses, and revisit intentions. As consumers frequently rely on online platforms such as Naver Place to assess cafés before visiting, the emotional tone of visual content, specifically images, plays a critical role in shaping perception. However, few studies have empirically investigated how the emotional characteristics of online spatial imagery relate to measurable customer responses. This study aimed to quantitatively examine the relationship between the emotional attributes of café space images and customer evaluation indicators, including star ratings and user-generated reviews. (Method) The study followed a three-step analysis. First, 5,548 representative images were collected from Naver Place, a leading Korean location-based platform. A VGG16-based convolutional neural network was used to classify and filter 3,428 spatial images, excluding non-spatial images such as food or signage. Second, OpenAI’s CLIP model was employed to extract multimodal image embeddings based on seven predefined emotional descriptors (warm and inviting, sleeve and clean, eco-friendly, and organic). K-Means clustering was then applied to categorize the images into the corresponding emotional types. Third, one-way ANOVA and Tukey’s HSD post-hoc tests were used to analyze the statistical relationship between emotional clusters and three customer evaluation metrics: star ratings, visitor review counts, and blog review counts. (Results) The results revealed distinct regional variations in the emotional image distribution and corresponding customer responses. For instance, cafés in Garosu-gil exhibiting Classic and Timeless or Warm and Inviting visual styles had significantly higher visitor review counts. In Samcheong-dong, the Sleek and Clean clusters were strongly associated with higher engagement. In Mangwon-dong, images classified as eco-friendly and organic or luxurious and elegant were correlated with elevated star ratings. Blog review counts also showed statistically significant differences depending on the emotional type, underscoring the effect of localized visual preferences on user behavior. (Conclusions) By integrating deep learning–based image classification with multimodal emotional embedding and statistical validation, this study provides empirical evidence that emotional image characteristics influence online customer evaluations. The findings offer a data-driven framework for understanding affective perception in spatial imagery and highlight the importance of emotion-sensitive design strategies in commercial spaces. The proposed methodology can be extended to other urban contexts and digital platforms for evidence-based space planning and customer engagement analysis.