南京理工大学学报（自然科学版）2025，Vol.49Issue(5)：600-611,624,13.DOI:10.14177/j.cnki.32-1397n.2025.49.05.009
不同天气条件下自然生长苹果的鲁棒性识别
Robust recognition of naturally grown apples under different weather conditions
摘要
Abstract
This paper proposes a robust recognition method based on domain adaptive object detection to address the interference of complex backgrounds and weather conditions in visual recognition for orchard-picking robots.Traditional methods require extensive data collection and annotation to enhance model performance,leading to high costs.The study constructs a hybrid dataset combining self-collected and open-source data,and generates degraded samples under rainy,foggy,and low-light conditions through image synthesis techniques.The proposed AppleDet model integrates real-time detection transformer(RT-DETR)with CLIP's multimodal architecture,developing contrastive language-image pre-training instance normalization(CLIPIN)domain adaptation through instance normalization for feature distribution transformation.A five-level chain domain adaptation strategy progressively optimizes weather adaptability,while random sampling augmentation simulates real domain characteristics.Experimental results demonstrate significant performance improvements:the AppleDet achieves an average precision(AP)of 94.0%in sunny conditions,with 78.8%,82.5%,and 90.9%AP under rainy,foggy,and low-light scenarios respectively,showing 10.4%,2.6%,4.6%enhancements over baseline models.The research proves that semantic-guided domain adaptation effectively mitigates weather-induced domain shifts,providing a reliable visual solution for all-weather orchard robotic operations.关键词
果园采摘机器人/深度学习/数据集构建/域适应目标检测Key words
orchard-picking robot/deep learning/dataset construction/domain adaptive object detection分类
计算机与自动化引用本文复制引用
郭许景,郭昊翔,陈壹风,孙建国,郭健..不同天气条件下自然生长苹果的鲁棒性识别[J].南京理工大学学报（自然科学版）,2025,49(5):600-611,624,13.基金项目
江苏省重点研发计划项目(BE2021016) （BE2021016）
