An automatic detection method for cervical liquid-based cells based on the improved YOLO V5s
Xudong Shen, Tao Yebo, WU Xianglian, Linfei CHEN, Shitao Shen · Research Square · 2023
Abstract To address the insufficient local inference performance of existing object detection algorithms used in cervical liquid-based cytology, we propose an enhanced YOLOv5s network structure. This architecture dynamically adjusts the weights of channels and spatial attention modules, thereby improving the extraction of feature information from small objects and enhancing the network model's detection capabilities. Alongside this enhanced network model, we studied the Mixup data augmentation technique, which effectively increases the sample size and addresses data imbalance in the custom dataset. We employ CIoU as the loss function for bounding box regression, thereby improving the localization accuracy of the network model's bounding boxes. Comparative experiments on a self-compiled cervical liquid-based cytology dataset show that the improved algorithm achieves a mean Average Precision (mAP) of 0.921, a 5.6% point increase compared to the original YOLOv5s, thus validating the effectiveness of our approach.