Мультимедійна система валідації об’єктів міської інфраструктури: порівняльний аналіз нейромережевих моделей та експертних оцінок

Abstract

This paper presents a hybrid multimedia validation system for crowdsourced urban accessibility reports that combines convolutional neural networks with a Human-in-the-Loop verification interface. Uploaded photos are first processed by CNN-based classifiers and perceptual-hashing to detect and de-duplicate candidate obstacles (e.g., ramps, high curbs, missing tactile tiles). Each prediction is accompanied by a confidence score; low-confidence or ambiguous cases are flagged as disputed and routed to users for manual review. User confirmations and corrections are fed back into an active-learning pipeline to incrementally improve model performance. The proposed two-stage architecture balances automated scalability with human judgment, minimizing erroneous automatic decisions in challenging visual conditions. We evaluate the approach on a representative crowdsourced dataset and discuss metrics for accuracy, annotation effort, and user validation time. Practical considerations for privacy, edge vs. cloud deployment, and UX design are addressed. The system aims to produce reliable, georeferenced accessibility maps to support safer route planning for mobility-restricted populations