A-Preprocessing and Labelling,
B-Data augmentation, C- Feature
extraction, D-Defects
classification recognition
Fig. 4 Flow chart of surface defect identification and classification of
Si3N4 ceramic bearing inner ring
⑴Preprocessing and labeling, the prerequisite for deep learning models
to obtain better recognition and classification results is labeling the
training set [20]. The contrast between the
defective part and the background is enhanced by image preprocessing.
LabelIMG is used to mark the surface defects, which are Pit, Crack, Wear
and Snowflake. The XML file corresponding to the defect is generated. It
is used to store the information of
Si3N4 ceramic bearing roller inner ring
surface defect, including the location, name, shape and boundary frame
information of the defect. The boundary frame contains the pixel
information of the surface defect.
⑵Data augmentation, large-scale dataset is beneficial to obtain better
recognition and classification results. Online data augmentation expands
the size of the dataset, improves the robustness of the target detection
model, prevents the occurrence of overfitting.
⑶Feature extraction, the expanded data set is input from the input layer
to the convolutional neural network and output from the output layer.
And it is subsequently passed through CONV layer, pooling layer, CONV
layer, pooling layer and fully connected layer. The images are
transformed into two-dimensional matrices. Image features are
preliminary extracted. Statistical features are extracted and image
features are recorded according to regions.
⑷Defects classification recognition, smaller feature matrix is got. The
features of each part are summarized. ReLU activation function is used
to output feature values. The process of surface defect image
recognition and classification has been completed.