基于卷积特征的光纤缺陷检测方法

Defect detection method for fiber based on convolutional neural network

针对人工难以实现熊猫型保偏光纤的缺陷检测问题,提出了一种小样本下基于卷积特征的图像检测方法。首先针对光纤缺陷特征将Inception V3模型微调,使用微调后模型提取光纤的2048维卷积特征;其次使用主元分析法将2048维特征降为74维;最后使用降维后的特征训练支持向量机分类器,同时使用粒子群算法对分类器参数寻优,实现对光纤缺陷的识别与分类。经实验证明,该方法对光纤涂覆层微小缺陷的识别率达到97%,涂覆层局部损伤和严重破损的识别率达到100%,对降低光纤环绕制中原纤损失、提升光纤环的精密性、研究光纤缺陷对光纤陀螺精度的影响有一定意义。

Because it is difficult to realize the defect detection of Panda-type polarization-maintaining fiber, a method based on feature of convolutional neural network in small set is proposed. First, the Inception V3 model is fine-tuned for the fiber defect feature, and the 2048-dimensional convolution feature of the fiber is extracted using the fine-tuned model. Secondly, using the principal component analysis method, the 2048-dimensional feature is reduced to 74 dimensions. Finally, the dimension-reduced features are input into training the support vector machine classification. At the same time, the particle swarm optimization algorithm is used to optimize the parameters of support vector machine to realize the identification and classification of fiber defects. The experimental results show that the recognition rate of the micro-defects of the fiber coating layer is 97%, and the recognition rate of local damage and severe damage of the coating layer reaches 100%. This algorithm has significance for reducing the loss of fibril in the fiber-wound system, improving the precision of the fiber-optic ring, and studying the influence of fiber defects on the accuracy of the fiber-optic gyroscope.