Research On Surface Defect Detection Algorithms Of Rare-Earth Magnetic Materials Based On Deep Learning

Nowadays,with the rapid development of modern manufacturing industry,manufacturing processes are developing towards automation and intelligence.As one of the important basic material in modern industrial,the rare-earth magnetic materials production is characterized by various complicated procedures.Improper operation of production equipment,mechanical failure or human negligence may cause defects in product appearance.With the increase of the production of rare-earth magnets,the manual detection speed can not meet the requirements of production.At the same time,traditional rare-earth enterprises rely on manual sorting,which will lead to an increase in labor cost.Therefore,we take rare-earth magnetic pieces as the detected object,and defect recognition algorithm based on deep learning is adopted for automatic detection.The specific research contents are as follow:(1)Lightweight residual neural network with multi-scale feature(LMSF-Res Net)is proposed on the classification of surface defects of magnetic pieces.This work aims to recognize common defects on the surface of workpiece and realizes high-accuracy real-time detection of different types of damage on the surface.We reduce the calculation and parameters of the neural network by using depthwise separable convolution and channel split operator,and merge features from different branches at distinct scales by using the structure of multi-scale model.Then,we employ the fusion strategy of channel attention mechanism to improve the prediction accuracy at the same time.Under the condition of a relatively low number of parameters,the experimental results show that the accuracy of the model is up to 98.75%,and the rate of model reasoning on the experimental platform is 0.082 s,which meets the requirements of real-time detection.(2)A defect detection method for magnetic pieces based on improved Single Shot Detector(SSD)was proposed.In order to meet the needs of different applications,we make a further study on the basis of defect image classification,and carry out object detection for several defect types with location information.Firstly,SSD is selected as the defect detection algorithm,but the result is not satisfactory of the algorithm in the detection of small defects.Then,we embedded the multi-scale receptive field module into the backbone network of the algorithm to improve the feature extraction ability of the model,and integrated the strategy of feature fusion of PANet(Path Aggregation Network)into the model.In order to enhance the detection ability of the model,the semantic information of high-level is strengthened by an efficient channel attention mechanism.The detection speed of the improved SSD algorithm is 55 FPS,and the m AP(mean Average Precision)is up to 83.65%,which is 3.41% higher than the original SSD algorithm,and the ability to identify small defects is significantly improved.