Individual identification of animals is crucial for understanding their behavior,ecology, and conservation, yet traditional methods are often invasive,labor-intensive, and prone to observer bias. Recent advancements in artificialintelligence (AI), particularly deep learning (DL) and computer vision(CV), offer non-invasive solutions. Existing AI-based primate facial recognitionsystems (AI-PFRS) are often limited by reliance on manual annotationand a strong frontal-view bias. This study addresses these gaps by developingand validating a scalable AI-PFRS for Japanese macaques (Macaca fuscata).Research was conducted on 33 focal individuals housed in a semi-free, 4-hectare forest enclosure, providing a naturalistic social and environmentalcontext. We introduce a novel, mostly-automated annotation pipeline thatleverages an isolation filming protocol and strict video naming conventions totransform generic YOLO-based face detections into per-identity annotatedtraining data with minimal manual effort. Applied across four seasonal datacollection periods (Summer 2023 through Winter 2025), this approach generatedover 720,000 annotated frames encompassing diverse poses, lighting, anddistances, critically capturing the high diversity in the environment due toseasonal changes and temporal variability in morphological features during mating and non-mating seasons. Model robustness was enhanced througha content-aware, difficulty-stratified train/validation split. Performance significantlyimproved over time, with the correct identification rate increasingfrom 57.60% to 93.30%. While validation performance was strong, accuracyon a more challenging test set that more closely reflects day-to-day field conditionswas 82.1%, consistent with real-world deployment complexities. Thiswork presents an ecologically grounded framework for non-invasive primateidentification, with direct implications for long-term behavioral monitoringand Citizen Science initiatives, fostering more effective and ethical wildlifemanagement strategies.