首页|优化Faster R-CNN算法的小样本缺陷检测研究

优化Faster R-CNN算法的小样本缺陷检测研究

扫码查看
随着自动化检测技术的发展,基于深度学习的缺陷检测技术以其高精度、高效率、非接触性的特点正逐渐成为工业和学术领域的研究热点.为解决实际工业生产中由于产品缺陷数据集样本不足、类别不均衡导致的模型过拟合、检测精度低等问题,提出了一种基于Faster R-CNN算法框架优化的缺陷检测模型Faster R-CNN-H-BFC,通过基于多层感知器(multi-layer perceptron,MLP)实现的幻觉网络能够从具有丰富样本的基类中学习到类共享特征并为新类生成额外的幻觉样本供模型训练,并且针对Faster R-CNN本身存在的识别精度低以及检测效果差等问题,将原始的VGG16主干网络替换为具有残差结构的ResNet50,并引入了特征金字塔网络(Feature Pyramid Networks,FPN)实现多尺度特征融合,添加混合注意力机制(Convolutional Block Attention Module,CBAM)来增强模型的特征提取能力.实验和数据表明:改进后的缺陷检测模型在极少样本场景下具有较好的检测效果,平均检测精度相较于改进前提升了 3.11%.
Research on small sample defect detection based on optimized Faster R-CNN algorithm
With the development of automatic detection technology,defect detection technology based on deep learning is gradually becoming a research hotspot in industrial and academic fields due to its high precision,high efficiency and non-contact characteristics.In order to solve the problems of model overfitting and low detection accuracy caused by insufficient samples and unbalanced categories of product defect data sets in actual industrial production,a defect detection model Faster R-CNN-H-BFC based on Faster R-CNN algorithm framework optimization is proposed.The illusion network based on multi-layer perceptron(MLP)can learn class sharing features from base classes with rich samples and generate additional illusion samples for new classes for model training.Aiming at the problems of low recognition accuracy and poor detection effect of Faster R-CNN,the original VGG16 backbone network is replaced with ResNet50 with residual structure,and Feature Pyramid Networks(FPN)is introduced to realize multi-scale feature fusion,and the Convolutional Block Attention Module(CBAM)is added to enhance the feature extraction ability of the model.Experiments and data show that the improved defect detection model has better detection effect in very few sample scenarios,and the average detection accu-racy is 3.11%higher than that before improvement.

small-sampleFaster R-CNNillusion networkfeature pyramid networkattention mechanismdefect detecting

何军红、温观发、黎长鑫

展开 >

西北工业大学航海学院,陕西西安 710072

小样本 Faster R-CNN 幻觉网络 特征金字塔网络 注意力机制 缺陷检测

2024

工业仪表与自动化装置
陕西鼓风机(集团)有限公司

工业仪表与自动化装置

CSTPCD
影响因子:0.393
ISSN:1000-0682
年,卷(期):2024.(5)