YOLOv8n-CSD:轻量化的金漆画题材检测模型OA
YOLOv8n-CSD:Lightweight Gilded Lacquer Painting Detection Model
金漆画作为重要的非物质文化遗产,对其题材的精准检测是数字化保护和传承的基础环节.然而,该类图像具有纹饰繁复、目标尺度多样等特点,大多数卷积检测算法固有的下采样与池化操作易造成对微小目标与复杂结构的感知能力不足.针对上述问题,本文基于YOLOv8n构建一种轻量级金漆画题材检测模型YOLOv8n-CSD.首先设计上下文特征增强模块,通过大感受野卷积、多分支结构与通道注意力机制,增强模型对上下文、纹理和语义信息的建模能力;其次,设计动态梯度感知检测头,通过动态调整专家响应机制,提升小目标、稀疏与密集目标及复杂目标的检测精度;最后,引入空间和通道协同注意力机制,提升模型对复杂特征的理解能力与深层信息的融合效果.实验结果表明,YOLOv8n-CSD在自建金漆画数据集上的mAP@0.5相较原YOLOv8n提升5.7百分点,在COCO数据集和DEArt数据集上的mAP@0.5相较原YOLOv8n分别提升3.3百分点和4.9百分点,实验结果验证了改进后的模型的有效性与泛化能力,说明该模型对艺术图像中复杂的纹理结构、稀疏目标分布等的检测具有更强的适应性.
Gilded lacquer painting are an important form of intangible cultural heritage,and accurate thematic detection of them serves as a fundamental step for their digital preservation and inheritance.However,such images are characterized by intri-cate decorative patterns and significant variations in target scales.The inherent down sampling and pooling operations in most convolution-based object detection algorithms often result in insufficient perception of small targets and complex structures.To address these challenges,this paper proposes a lightweight thematic detection model for lacquer paintings,termed YOLOv8n-CSD,based on the YOLOv8n framework.First,a context feature enhancement module is designed to improve the modeling of contextual,textural,and semantic information by integrating large-receptive-field convolutions,a multi-branch architecture,and a channel attention mechanism.Second,a dynamic gradient-aware detection head is developed to enhance the detection ac-curacy of small,sparse,dense,and structurally complex targets through an adaptive expert response mechanism.Finally,a spatial-channel collaborative attention mechanism is introduced to strengthen the model's capability for complex feature repre-sentation and deep feature fusion.Experimental results demonstrate that YOLOv8n-CSD improves the mAP@0.5 by 5.7 percent-age points over the original YOLOv8n on the self-built gilded lacquer painting dataset.Furthermore,it achieves gains of 3.3 and 4.9 percentage points in mAP@0.5 on the COCO and DEArt datasets,respectively.These results validate the effectiveness and generalization capability of the proposed improvements,indicating that the enhanced model exhibits stronger adaptability in de-tecting artistic images,including complex texture structures and sparse target distributions.
缪远;张豪然;邢树礼;凌京辉;梁浩;梁艳
福建理工大学设计学院,福建 福州 350100福建理工大学计算机科学与数学学院,福建 福州 350100福建理工大学计算机科学与数学学院,福建 福州 350100福建理工大学计算机科学与数学学院,福建 福州 350100福建理工大学计算机科学与数学学院,福建 福州 350100福建理工大学互联网经贸学院,福建 福州 350100
信息技术与安全科学
YOLOv8n金漆画上下文特征增强SCSA动态梯度感知检测头
YOLOv8ngilded lacquer paintingcontext feature enhancement moduleSCSAdynamic gradient sensing detec-tion head
《计算机与现代化》 2026 (9)
65-76,12
国家社科基金资助项目(24BG148)
评论