Research On Online Inspection Technology Of Magnetic Tiles Based On Deep Learning

Tile-shaped permanent magnet ferrite(magnetic tile)is an important component of the motor,and the surface defects of the magnetic tile will directly affect the performance and life of the motor,so the detection of surface defects on magnetic tiles is of great significance.With the successful application of deep learning technology in the field of computer vision,deep learning-based methods are gradually applied to the surface defect detection of magnetic tiles.The current methods are all supervised learning methods that require large amounts of labeled data to train deep neural networks,which are laborious and time-consuming for the dataset construction.Therefore,these methods are difficult to adapt to real-world cases where product characteristics change due to the change of market demand or the optimization of production.In this paper,a semi-supervised deep learning method was proposed for the surface defect detection of magnetic tiles,aiming to reduce the need of labeled data while exploiting the powerful representation and learning abilities of deep neural networks.The main contents of this paper are as follows:(1)The design and construction of an online image acquisition system for magnetic tiles.This paper designed and built an online image acquisition system for magnetic tiles,which uses an STM32 microprocessor to control the transmission and positioning system,and a Raspberry Pi to control the camera to capture images of six sides of magnetic tiles,and uses diffused lighting to achieve uniform exposure.The collected images were then labeled with defects to create a surface defect detectiondataset.(2)The proposal of a semi-supervised surface defect detection algorithm based on pseudo-labels for magnetic tiles.The proposed method contains two models:a teacher model and a student model.The training process is divided into two stages:pseudo-label generation and student model training.The pseudo-label generation stage alternatively optimizes the teacher model and pseudo-labels of unlabeled data based on the idea of transductive learning to obtain high-quality pseudo-labels.The student model training stage uses the generated pseudo-labels to help the training of the defect detection model.The experimental results showed that with only 4.4%of labeled samples in the training set,the proposed method was able to achieve 90.13%defect detection accuracy,which is 5.42%higher than the supervised method.(3)The compression and deployment of the defect detection model.The proposed method replaced the student model with a lightweight network based on the idea of knowledge distillation,while the teacher model still using a complex network,thus enabling the compression of the defect detection model without significantly degrading the model performance.The compressed model was deployed on Raspberry Pi using the deep learning inference framework MNN,and the experimental results showed that the inference time of the compressed model on Raspberry Pi is only 42 ms,which can meet the requirement of real-time performance.