Abstract: Accurate detection of small foreign fibers mixed in cotton is the basis and key to guarantee the quality of yarn and fabric.Aiming at the problems of high leakage rate and complex network structure of existing algorithms in the detection of small foreign fibers in cotton,an improved Single Shot multibox Detector (SSD) based on convolutional neural network lightweight was proposed for the detection of foreign fibers in cotton.The original backbone feature extraction network VGGNet16 in the SSD algorithm was replaced with MobileNetv2 network by introducing innovative designs such as depth-separable convolution and inverted residual structure;for the problem that the candidate box size generated in SSD algorithm did not match the size of cotton foreign fibers leading to a high percentage of the cotton background,which caused the imbalance of the positive and negative samples,the K-means++ algorithm was used to determine the candidate box size of cotton foreign fibers and make cluster analysis,thus the cotton foreign fiber size and the candidate frame size were corrected  according to the clustering results.The results showed that the proposed method effectively improved the effect of foreign fiber detection and computational efficiency while realizing model lightweighting.

Key words: foreign fiber detection, improved single shot multibox detector, convolutional neural network, K-means++clustering, lightweighting

摘要： 精准检测棉花中混杂的小型异纤是保障纱线与织物质量的基础和关键。针对现有算法在棉花小型异纤检测中存在的漏检率高、网络结构复杂等问题,提出一种基于卷积神经网络轻量化的改进单步多框检测器(SSD)的棉花异纤检测方法。首先,通过引入深度可分离卷积、倒残差结构等创新性设计,将SSD算法中原有骨干特征提取网络VGGNet16替换为MobileNetv2网络;然后,对于SSD算法中生成的候选框尺寸与棉花异纤大小不匹配导致棉花背景占比过高,从而引起正负样本不均衡的问题,采用K-means++算法对棉花异纤尺寸进行聚类分析,根据聚类结果修正候选框尺寸。通过算例进行验证,结果显示所提方法在实现模型轻量化的同时有效提升了异纤检测效果和计算效率。

关键词: 异纤检测, 改进SSD, 卷积神经网络, K-means++聚类, 轻量化