基于YOLOv8改进的服装疵点检测算法  

Improved garment defect detection algorithm based on YOLOv8

作　　者：鲍禹辰 徐增波[1] 田丙强[1] BAO Yuchen;XU Zengbo;TIAN Bingqiang(School of Textiles and Fashion,Shanghai University of Engineering Science,Shanghai,China)

机构地区：[1]上海工程技术大学纺织服装学院,上海

出　　处：《东华大学学报（自然科学版）》2024年第4期49-56,共8页Journal of Donghua University(Natural Science)

摘　　要：针对服装疵点检测方法,提出了基于YOLOv8改进的算法YOLOv8-MBRGA,用于完成服装疵点的检测任务。引入BiFPN金字塔替换head层中的concat连接,将语义信息传递到不同的特征尺度上,从而增强特征融合。为加速模型的收敛速度和推理速度,在检测头上增加RepVGG网络,有助于更好地训练深层次的网络模型。采用分离卷积替换Conv卷积降低网络的复杂度并融入注意力机制EffectiveSE增强模型的特征提取和多尺度信息融合的能力。试验结果表明,YOLOv8-MBRGA算法在服装疵点检测上获得了显著的效果,平均精度均值提高了5.50%,精确度提高11.06%,在推理速度基本保持不变的情况下,模型的计算量下降30.48%。An improved algorithm based on YOLOv8,YOLOv8-MBRGA,is proposed for the task of detecting garment defects.The BiFPN pyramid is introduced to replace the concatenation in the head layer to transfer semantic information to different feature scales,thus enhancing feature fusion.To accelerate the convergence speed and inference speed of the model,RepVGG network is added to the detection head,which helps to better train the deep network model.Separate convolution is used to replace Conv convolution to reduce the complexity of the network and incorporate the attention mechanism EffectiveSE to enhance the feature extraction and multi-scale information fusion of the model.The experimental results show that the YOLOv8-MBRGA algorithm obtains significant results in garment blemish detection,with an increase of 5.5% in the mean average accuracy,an increase of 11.06%in precision,and a reduction of 30.48% in the computational effort of the model,while the inference speed remains essentially unchanged.

关 键 词：服装疵点 BiFPN金字塔 RepVGG网络 YOLOv8 

分 类 号：TS941.26[轻工技术与工程—服装设计与工程]