ConvNeXt knowledge distillation optimization method for SPOTS-10 animal pattern recognition

• 발행기관: 한국컴퓨터게임학회
• 저자명: Dae-Won Park Changyu AO Seung-Eon Jeong Soo-Kyung Moon Youn-Mo Soung Man-Sung Kwen Uk Cho Dae-In Kang Sung-Ho Jung Gwang-Jun Kim
• 간행물 정보: 『한국컴퓨터게임학회논문지』제38권 1호, 88~97쪽, 전체 10쪽
• 주제분류: 공학 > 컴퓨터학
• 파일형태: PDF
• 발행일자: 2025.03.31

국문 초록

Animal pattern recognition from nighttime grayscale images is crucial for wildlife protection and ecological monitoring. Most of the current models suffer from a large parameter scale, making them unsuitable for deployment in resource-constrained environments. To address this challenge, this study proposes a multi-layer knowledge distillation approach based on the teacher Convnext model to improve the lightweight student model's classification performance effectively. Experimental results show that the parameters of the distilled CifarResNet20 model are only 0.27M, and the accuracy is 88.76%, which is superior to the traditional single-layer distillation and another tiny student model. The study confirms the efficiency and practical value of the proposed method in practical applications such as ecological monitoring.

목차

1. Introduction

2. Related Work

3. Materials and Methods

4. Result

5. Conclusion

참고문헌

키워드

Animal Pattern Recognition Knowledge Distillation ConvNeXt Lightweight Model