人工智能目标检测技术在书画文物病害调查中的应用  

Application of artificial intelligence object detection technology in disease identification of calligraphy and painting cultural relics

作　　者：邓旭帅 李子璇 张云春[2] 沐蕊 DENG Xushuai;LI Zixuan;ZHANG Yunchun;MU Rui(School of History and Archives,Yunnan University,Kunming 650091,China;School of Software,Yunnan University,Kunming 650504,China;Yunnan Provincial Museum,Kunming 650214,China)

机构地区：[1]云南大学历史与档案学院,云南昆明650091 [2]云南大学软件学院,云南昆明650504 [3]云南省博物馆,云南昆明650214

出　　处：《西北大学学报(自然科学版)》2025年第1期98-105,共8页Journal of Northwest University(Natural Science Edition)

基　　金：国家档案局科技项目计划(2021-B-03、2022-B-002)。

摘　　要：针对书画文物保护工作中人工病害调查和病害图绘制效率低的问题,探索了基于深度神经网络的目标检测技术识别书画病害的可行性。选择YOLOv5系列模型并根据本研究任务特点对其结构做了优化,包括FGSM算法、CmBN策略、Dropblock正则化和CIOU-Loss损失函数。利用博物馆馆藏书画文物素材,融合Mosaic数据增强方法进行书画文物图片的增强,设计了滑动窗口检测技术、图像逐层分析和定位裁剪技术,初步训练出了2个具备病害识别功能的模型,根据模型性能检验指标最终选择了YOLOv5x6作为本研究任务的模型。测试结果表明,该模型以较高的准确率和查全率识别出了待检测病害,用时仅为人工的千分之一。该技术的引入可极大提高文物病害识别效率,并且在病害识别过程中保持客观、稳定的标准。Targeting the low-efficiency problem of manual disease identification and disease mapping in the protection of calligraphy and painting cultural relics,this paper explores the feasibility of deep neural network-based object detection technology to identify calligraphy and painting diseases.We design a series of YOLOv5 models with some architecture optimizations based on the special requirements of disease identification.The optimizations include FGSM algorithm,CmBN strategy,Dropblock normalization,and CIOU-Loss loss function.Using the materials of calligraphy and painting of cultural relics in the museum as inputs,we enhance the images by combining Mosaic data enhancement method.Two disease identification deep learning models are trained based on some improvements,including sliding-window detection,image clipping based on layer-by-layer image analysis and positioning,etc.By evaluating the models with bench-mark performance metrics,this paper chooses YOLOv5x6 for our task.The experimental results show that YOLOv5x6 outperforms the other models with the best precision and recall.This model takes one-thousandth time compared with manual work.The introduction of deep learning techniques in disease identification not only helps to improve the efficiency of disease identification of cultural relics,but also provides objective and stable standards in the processes of disease identification.

关 键 词：病害识别 目标检测 图像处理 YOLOv5 深度神经网络 

分 类 号：K854.3[历史地理—考古学及博物馆学]