A Color Attribute–Organization–Perception (CAOP) Framework for Assessing Urban Building Color Perception Based on Street-View Images: A Case Study of Shenyang, China

Urban building color is a key environmental factor shaping visual identity and public spatial experience. However, the relationships among basic color attributes, objective color organization, and public perception remain poorly integrated, and the applicability of cross-regional perception models to local settings requires clearer boundaries. We propose a Color Attribute–Organization–Perception (CAOP) framework that evaluates building color through three progressive layers: basic attributes, objective organization, and public perception. Using Shenyang, China, as a case study, we collected 56,520 street-view images and applied a SegFormer model with ADE20K-trained weights to extract building façades. Evaluation against locally annotated samples yielded a building-class IoU of 0.75, an F1-score of 0.82, and an overall mIoU of 0.71. After quality control, 38,274 valid samples entered statistical analysis and K-means clustering. For cluster-number selection, the elbow method indicated approximately K = 4, while the silhouette coefficient and Calinski–Harabasz index supported K = 2. The Davies–Bouldin index favored larger K values, and equal-weight rank aggregation produced a tie between K = 2 and K = 5. Considering model parsimony and repeated-subsample stability (ARI = 0.944 ± 0.026), we used K = 2 as a coarse-grained typology. Building color in Shenyang showed pronounced spatial heterogeneity and road dependence. The central urban area generally had higher color complexity, effective number of colors, and contrast. Peripheral areas were more stable, although local hotspots remained near transport corridors and functional nodes. Lightness dispersion was strongly associated with color contrast, while dominant color share was positively associated with hierarchy clarity and negatively associated with complexity and effective number of colors. Clustering identified two patterns, “dominant-color control–clear hierarchy” and “multicolor composition–continuous richness”, accounting for 24.8% and 75.2% of the samples, respectively. The absolute Cliff’s δ values for all six perception indicators were below 0.08, and the difference in perceived liveliness was not significant (q = 0.396). These indicators were inferred from images by a Place Pulse 2.0-derived model and were used only for within-sample comparisons. They do not represent direct evaluations by Shenyang residents. The study establishes a continuous analytical pathway from machine-recognized basic color information and objective color organization to model-inferred perception. It supports quantitative diagnosis of urban building color and preliminary screening of candidate street segments, while specific renewal strategies require field investigation and local public evaluation.