基于预训练模型的深度学习算法及其在图书馆行人目标检测中的应用  

Deep Learning Algorithm Based on Pre-trained Models and Its Application in Pedestrian Target Detection in Libraries

作　　者：严珊 Yan Shan

机构地区：[1]中南财经政法大学图书馆,湖北武汉430073

出　　处：《图书馆研究与工作》2024年第3期43-51,共9页Library Science Research & Work

摘　　要：图书馆行人目标检测能够实现对图书馆内行人目标情况的统计,观察读者的学习行为和时间倾向,对提高服务质量和改善图书馆设施构造具有重要作用。现有图书馆行人目标深度学习算法能够对行人目标进行自动识别和统计,但计算复杂度高,神经网络模型的训练效率低,难以适应图书馆不同场所的需求。对上述问题,文章提出一种基于预训练模型的深度学习算法。该算法基于迁移学习的思想,对模型进行预训练,从而避免模型从零开始训练,并且设计了一种广义损失函数,该函数不仅关注不同对象的重合区,还关注不重合区,从而能更好地体现出两个对象的重合性。实验结果表明,基于预训练模型的深度学习算法能够提高行人目标检测模型的训练效率以及检测的精确度和查全率,能够满足图书馆不同场景下行人目标检测的需求。Reader detection in libraries enables the statistical analysis of reader activity,observation of readers'learning behaviors,and identification of time trends,playing a crucial role in enhancing service quality and improving library facility design.Currently,manual video monitoring is the primary method for reader detection in libraries,which is timeconsuming,labor-intensive,and lacks accuracy in counting readers.Existing deep learning algorithms for library reader detection can automatically identify and count readers,but they exhibit high computational complexity,low efficiency in training neural network models,and difficulty in adapting to the diverse requirements of different library settings.Moreover,occlusion between shelves and furniture in libraries disrupts the structural information of readers,leading to potential omission errors.To address these challenges,this paper proposes a deep learning algorithm based on pre-trained models.The algorithm,inspired by transfer learning,pre-trains the model to avoid training from scratch.Additionally,a generalized loss function is designed,focusing not only on the overlapping regions of different objects but also on nonoverlapping regions,better reflecting the overlap between two objects.Experimental results demonstrate that the proposed method improves the training efficiency of reader detection models,enhances detection accuracy,and achieves satisfactory recall rates in various library scenarios.

关 键 词：行人目标检测 深度学习算法 YOLOv3检测算法 预训练模型 图书馆 

分 类 号：G250.7[文化科学—图书馆学]