基于多流神经网络的PET瓶坯缺陷检测研究

Research on PET Preform Defect Detectionbased on Multi-Stream NeuralNetwork

首次提出了一种基于多流卷积神经网络PET瓶坯缺陷分类和识别系统。近年来,基于多流网络的特征融合方法在缺陷分类与质量检测中具有良好的应用前景。设计PET瓶坯质量检测平台用来采集图像样本,通过图像增强方式扩展样本数量,以防止深度学习中的过拟合现象。利用Sobel算子计算图像的梯度,并对其进行归一化处理。设计了一种采用最大值融合策略的多流卷积神经网络模型,实现了relu5层的特征融合。上述模型以PET瓶坯的原始图像以及其对应的梯度图像和缺陷区域图像作为输入,通过子网络提取特征,实现特征融合,然后输入到SVM分类器进行缺陷分类。通过实验分析,证明上述模型具有良好的收敛性、准确性、稳定性和泛化能力。

In this paper, a defect classification and identification system of PET preform based on multi-stream convolutional neural network is proposed for the first time. Recent feature fusion methods based on multi-stream network prove promising performance for defects classification and quality prediction. The quality detection platform of PET preform was designed to collect image samples, and the number of samples was expanded through image enhancement to prevent over fitting in deep learning. The gradient of image was calculated using Sobel operator and normalized. A multi-stream convolutional neural network model adopting Maximum fusion strategy was designed to realize feature fusion on ReLU5 layer. The model takes the PET preform original image and its corresponding gradient image and defect region of interest as inputs, and extracts features through three sub networks to realize feature fusion, and inputs them to SVM classifier for defect classification. The experimental results show that the model has good convergence, accuracy, stability and generalization ability.