Abstract

Object detection models, such as the You Only Look Once (YOLO) family, have been widely deployed in agriculture, surveillance, and industrial inspection. However, selecting an appropriate model remains challenging due to trade-offs between precision, recall, and efficiency. Conventional evaluation approaches often rely on a single metric such as an F1-score, which may obscure critical domain-specific requirements. This study proposes a metrics-based framework that integrates multiple evaluation indicators into a unified composite index using Multi-Criteria Decision-Making (MCDM) principles. Precision, recall, and F1-score were normalized and weighted according to use-case priorities across four experimental scenarios: indoor/non-molted, indoor/molted, outdoor/non-molted, and outdoor/molted. Composite scores and rankings were computed under three operational contexts: security/surveillance, real-time edge applications, and quality inspection. Results show that Scenario 3 (outdoor/non-molted) consistently achieved the highest composite performance across all contexts, while Scenario 2 (indoor/molted) and Scenario 4 (outdoor/molted) varied in ranking depending on use-case weights. Sensitivity analysis further confirmed the robustness of Scenario 3 under shifting recall weight assignments. Compared to single-metric evaluation, the proposed framework offered finer-grained differentiation, highlighting trade-offs that are critical for real-world deployment. The findings demonstrate the value of multi-metric integration for systematic model selection and provide practical guidance for applying object detection in diverse operational environments.