HSDet: A Representative Sampling Based Object Detector in Cervical Cancer Cell Images

Abstract

Automated detection of cervical cancer cells has the potential to reduce error and increase productivity in cervical cancer screening. However, the existing object detection methods to detect the cervical cancer cells make inadequate use of negative samples and fail to balance the number of hard samples and easy samples. In this work, we propose a image-level sampling method, called pair sampling, which can extract the representative negative samples and generate the sample pair image. Then, based on the sample pair image, we propose the hybrid sampling, a proposal-level sampling method, to balance the number of hard samples and easy samples. Combining the proposed sampling methods, we design a representative sampling based cervical cell detector, which can detect the cervical cancer cells effectively. In order to comprehensively evaluate our method, we use a dataset which is consisted of 16000 annotated cervical cell images with size of \(1024\times 1024\). The experimental result shows that our method achieves 57.1% mean Average Precision (mAP), and it is higher than cascade R-CNN and Faster R-CNN by 4.7% and 5.8%, respectively.