ELARKHNU browse.title.page

Зараз показуємо 1 - 1 з 1

Efficiency Analysis of Wrecking Waste Classification Using Neural Network
(2025) Zalutska, O.; Mazurets, O.; Molchanova, M.; Мазурець, Олександр Вікторович
This study examines the efficiency of a neural approach to identifying and classifying wrecking waste in photographic data. Building on a two-factor design that couples scene-level detection with per-class binary refinement, we quantify where accuracy is earned, where it is lost, and which components: architecture, data composition, augmentation, and arbitration – most strongly govern the outcome. Using a composite dataset of ten material classes, we report that the hybrid method attains up to 97.8% classification accuracy, with substantial gains over a detector-only baseline on heterogenous or texture-confusable categories. Detector metrics such as mAP50 ≈ 0.746 and mAP50–95 ≈ 0.669 confirm reliable localization, while binary residual classifiers close the gap in label assignment, lifting macro F1 from ≈0.751 to ≈0.974. These findings indicate that efficiency, measured as correct material routing per unit inference cost, is maximized when detection and classification are decoupled yet reconciled through a calibrated arbitration rule.