基于双主干特征融合改进的服装图像分类识别算法  

Improved clothing image classification and recognition algorithm based on dual trunk feature fusion

作　　者：鲍禹辰 徐心放 余承志 BAO Yuchen;XU Xinfang;YU Chenzhi(School of Textiles and Fashion,Shanghai University of Engineering,Shanghai 201620,China;School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi′an,Shaanxi 710016,China)

机构地区：[1]上海工程技术大学纺织服装学院,上海201620 [2]陕西科技大学电智学院,陕西西安710016

出　　处：《毛纺科技》2025年第1期111-119,共9页Wool Textile Journal

摘　　要：针对复杂背景图像的服装分类与识别问题,提出YOLOv8、YOLOv10、YOLOv11主干的改进方案,通过2个并行的主干网络并结合BiFPN双向加权金字塔,实现特征信息的深度交互与高效融合。提出通过CSL注意力机制将DSConv、GhostConv、空间信息、通道信息与选择特征进行有效整合提高算法对服装特征信息的提取能力。通过deepfasion2数据集验证表明,改进方案YOLOv8Plus的精确度提高6.9%、召回率提高4.44%、平均精度mAP值提高5.86%。使用注意力机制CA、CBAM、CSL与YOLOv8结合并进行对比,CSL、CBAM、CA注意力机制mAP值分别高出基础YOLOv8算法1.75%、0.68%、0.78%。YOLOv8Plus方案相比YOLOv9c算法GLOPS下降74.04%,且mAP值仅降低0.39%。基于YOLOv10的改进方案YOLOv10Plus在精确度、召回率、mAP值上分别提高3.3%、0.45%、1.96%,基于YOLOv11的改进方案YOLOv11Plus在精确度、召回率、mAP值上分别提高2.12%、3.24%、4.47%,并且YOLOv10Plus与YOLOv11Plus的计算量分别是YOLOv9c的23.07%、22.11%,是RTDETR-ResNet101的16.89%、16.19%。结果表明,该研究方案在YOLO系列算法中mAP值具有有效的提高作用,相比于基础模型在服装分类识别应用上具有更优的效果。Aiming at the problem of clothing classification and recognition of complex background images,the improvement scheme of YOLOv8,YOLOv10 and YOLOv11 backbones was proposed to achieve the deep interaction and efficient fusion of feature information through two parallel backbone networks and combined with BiFPN bidirectional weighted pyramid.It was proposed to effectively integrate DSConv,GhostConv,spatial information,channel information and selective features through the CSL attention mechanism to improve the algorithm′s ability of extracting clothing feature information.The validation of the deepfasion2 dataset shows that the improved scheme YOLOv8Plus improves the precision by 6.9%,recall by 4.44%,and mAP value by 5.86%.Using the attention mechanisms CA,CBAM,and CSL in combination with YOLOv8 and comparing them,the CSL,CBAM,and CA attention mechanisms mAP value outperform the base YOLOv8 algorithm by 1.75%,0.68%,and 0.78%,respectively.The YOLOv8Plus scheme reduces the GLOPS of the algorithm compared to YOLOv9c by 74.04%and reduces the mAP value by only 0.39%.YOLOv10Plus,an improved scheme based on YOLOv10,improves precision,recall,and mAP by 3.3%,0.45%,and 1.96%,respectively,and YOLOv11Plus,an improved scheme based on YOLOv11,improves precision,recall,and mAP by 2.12%,3.24%,and 4.47%,respectively,and YOLOv10Plus and YOLOv11Plus are 23.07%and 22.11%of YOLOv9c and 16.89%and 16.19%of RTDETR-ResNet101 in computation,respectively.The results indicate that this research scheme has an effective improvement effect on mAP value in YOLO series algorithms,and has better performance in clothing classification and recognition applications compared to the basic model.

关 键 词：BiFPN CSL 服装分类与识别 YOLOv8 YOLOv9c YOLOv10 YOLOv11 

分 类 号：TS941.26[轻工技术与工程—服装设计与工程]