DeepLabv3++:一种基于语义分割的布匹瑕疵检测模型

DeepLabv3++:Fabric defect detection model based on semantic segmentation

摘要:

针对布匹瑕疵检测任务中存在的细小目标识别准确率低、检测速度慢等问题,提出一种新型的DeepLabv3++布匹瑕疵检测模型。首先,设计了多尺度的轻量级主干网络,用于提取形状各异、大小不一的瑕疵特征；其次,引入卷积注意力模块和空间通道注意力模块,分别实现对细小目标边界信息的捕捉和瑕疵区域的关注；接着,在解码部分添加两类多层次特征融合(multi-level feature fusion,MFF) 模块,以减少细节信息丢失问题；最后,采用工业现场采集的布匹瑕疵数据集对模型进行训练与评估。实验结果表明,本文提出的DeepLabv3++模型相较其他主流语义分割模型具有 更高的精度和检测速度,其参数量仅为410万,平均交并比(mean intersection over union,mIoU) 、平均像素准确率(mean pixel accuracy,mPA) 分别达到了90.01%、95.05%,满足了工业现场对检测精度和速度的均衡性需求。

Abstract:

A novel DeepLabv3++ model is proposed to address the low accuracy in identifying small targets and slow detection speed in fabric defect detection tasks.Firstly,a multi-scale lightweight backbone network is designed to extract features from defects with various shapes and sizes.Secondly,convolutional attention modules and channel spatial attention modules are introduced to capture boundary information of small targets and focus on defect regions.Additionally,two types of multi-level feature fusion (MFF) modules are added to mitigate the issue of detail information loss in decoder.Finally,the model is trained and evaluated using a fabric defect dataset collected from an industrial site.The results show that our DeepLabv3++ model outperforms other models,utilizing only 4.1 million parameters.It achieves a mean intersection over union (mIoU) of 90.01% and a mean pixel accuracy () of 95.05%,meeting the industrial site requirements for balancing detection precision and processing speed.