Title: Pixel-Based Icon Semantics: On-Device Visual Recognition for Complete iOS Icon Labeling Abstract: This paper presents an on-device visual recognition pipeline that achieves near-complete semantic labeling of icons in iOS applications using only screen pixels as input. By analyzing 327,879 icons extracted from iPhone app screens, we characterize a highly skewed distribution: 98 common icon types cover 92.8% of icons, while 331 long-tail types—many crucial to app functionality—account for the remaining 7.2%. To address this long-tail challenge, we introduce a hybrid classification system that combines a high-accuracy image classifier for common icons (96.3% accuracy) with a few-shot learning model for long-tail types (78.6% accuracy). The system further augments labels by detecting contextual nearby text (95.3% accuracy) and identifying modifier symbols within icons (87.4% accuracy), all without reliance on accessibility metadata or view hierarchies. Implemented entirely on-device, our approach supports 99.5% of collected icons—a significant increase over prior work—and validates label usefulness through a study with 23 participants. This work demonstrates a scalable, pixel-only perception system for real-time icon understanding on iOS, enabling applications in accessibility, UI search, and conversational agents.