Triplet Loss based Metric Learning for Cervical Cell Classification

Cervical cancer screening mainly relies on manual reading, which is time-consuming and laborious. To classify cervical cells more accurately and robustly, the triplet loss based metric learning has been proposed for cervical cell classification. First, Herlev data set is used to train the metric learning model with the bench hard triplet loss. The cervical cell images been mapped to the feature space by trained model and the corresponding feature vectors can be obtained. Second, the feature vectors and the cell labels are used to train the SVM, the trained SVM is used as a classifier for cervical cell diagnosis. The metric learning is utilized to extract features of cervical cells, and the SVM has been applied to distinguish cells. Feature visualization shows the feature expression of cells, and cells in same class have similar feature expression. Dimensionality reduction visualization illustrates that different cells have unique cluster centers and robust boundaries, which proves the reliability of feature extraction. Finally, the experimental results shown the reliability of feature extraction and the classification accuracy has also reached the state of the art. It is of great significance to the screening of cervical cells and can be applied to various cell diagnosis in the future.