摘要
在竹片表面缺陷检测中,竹片表面缺陷形状各异,成像环境脏乱,现有基于卷积神经网络(CNN)的目标检测方法面对这样特定的数据时检测准确率较低;竹片来源复杂且有其他条件限制,例如不同季节成色各异等限制,无法采集所有类型的数据,导致竹片表面缺陷数据量少,以至于CNN不能充分学习.针对以上问题,提出一种改进的ResNet50网络与迁移学习结合的竹片缺陷识别方法.首先,将获得的正负样本数据按照一定比例分为训练集、验证集和测试集;其次,利用OTSU算法和LBP算法对竹片图像进行二值化处理和特征提取,以减少噪音影响;最后,将ResNet50作为骨干网络加入L_(2)正则化和标签平滑与迁移学习结合,得到适应于竹片缺陷检测识别的优化模型.将所提检测网络与VGG16、DenseNet121、ResNet50以及目前常用于工业检测的YOLOv3分别在相同比例训练测试集上进行训练和测试.实验结果表明,在竹片数据集上所提检测网络的平均精度均值(mAP)比VGG16、DenseNet121、YOLOv3和ResNet50分别提高了23.45%、18.6%、19.51%和2.76%.所提方法能够针对形状各异的竹片表面缺陷进行有效检测,且降低时间消耗,在实际工业运用中具有很好的效果.
In the detection of bamboo surface defects,bamboo surface defects have different shapes and the imaging environment is messy.The existing target detection methods based on convolutional neural networks(CNNs)have low detection accuracy when facing such specific data.Moreover,the source of bamboo chips is complex and there are other restrictions,such as different colors in different seasons,so it is impossible to collect all kinds of data,resulting in a small amount of data on bamboo chip surface defects,so that the CNN can not fully learn.In view of the above problems,this paper proposes an improved bamboo defect identification method combining the ResNet50 network and transfer learning.Firstly,the obtained positive and negative sample data are divided into a training set,a verification set and a test set according to a certain proportion.Secondly,the OTSU algorithm and LBP algorithm are used to binarize the bamboo image and extract its features to reduce the influence of noise.Finally,the ResNet50 is used as the backbone network,and L_(2) regularization,label smoothing and migration learning are combined to obtain an optimized model suitable for bamboo defect detection and recognition.The proposed detection networks,VGG16,DenseNet121,ResNet50 and YOLOv3,which are commonly used in industrial detection at present,are trained and tested on the same scale training test set respectively.The experimental results show that the average precision mAP of the proposed detection networks is 23.45,18.6,19.51 and 2.76 percentage points higher than that of VGG16,DenseNet121,YOLOv3 and ResNet50,respectively.The proposed method can effectively detect the surface defects of bamboo chips with different shapes,and reduce the time consumption,which has a good effect in practical industrial application.
作者
常青
郑世祺
邓宇书
唐小琦
CHANG Qing;ZHENG Shi-qi;DENG Yu-shu;TANG Xiao-qi(School of Automation,China University of Geosciences,Wuhan 430074,China;Shaoyang Advanced Manufacturing Technology Research Institute,Shaoyang 422100,China;School of Mechanical Science&Technology,Huazhong University of Science and Technology,Wuhan 430074,China)
出处
《控制与决策》
北大核心
2025年第2期432-440,共9页
Control and Decision
基金
国家自然科学基金项目(52375520)
湖南省区域联合基金项目(2023JJ50037)
湖北省重点研发计划项目(2023BAB172,2023DJC173)。
