This study presents the development of an AI-based real-time on-device segmentation system designed to support recyclable waste sorting. A lightweight semantic segmentation model was implemented by combining the MobileViT-x-small backbone with the DeepLabV3 architecture, enabling pixel-level classification of recyclable items and intuitive visualization on a smartphone screen. A total of 200 real-world images were collected, with 150 used for training and 50 for testing. To enhance generalization under limited data conditions, the training set was expanded to 750 images through geometric and color-based augmentation techniques. The trained model was subsequently converted into ONNX format and deployed within a Flutter-based mobile application, allowing real-time inference directly on the device without reliance on external servers. The proposed system overlays semi-transparent masks and class labels onto the live camera feed, thereby reducing sorting errors and promoting active user participation in everyday recycling practices.

1. 서 론
2. 관련 연구
    2.1 Semantic Segmentation
    2.2 MobileViT
3. 시스템 설계
    3.1 데이터 수집
    3.2 데이터 증강
4. 모 델
    4.1 DeepLabV3-MobileViT-x-small 개요
    4.2 모델 구조
    4.3 손실 함수
    4.4 평가 지표
    4.5 학습 설정
    4.6 실험 환경 및 추론 성능
5. 결 과
    5.1 데이터 증강 전후 성능 비교
    5.2 최종 모델 선정
    5.3 클래스별 성능
6. 애플리케이션 구현
7. 결 론
    7.1 연구의 의의
    7.2 한계점
    7.3 향후 연구 및 발전 방향
    7.4 모델 경량화
    7.5 종합적 결론
References

• Min-Jun Kim(Department of Software and Communication Engineering, Hongik University) | 김민준 (홍익대학교 소프트웨어융합학과)
• Chunghun Ha(Department of Industrial and Data Engineering, Hongik University) | 하정훈 (홍익대학교 산업․데이터공학과) Corresponding author