DeepSperm: A robust and real-time bull sperm-cell detection in densely populated semen videos

Object detection is a primary research interest in computer vision. Sperm-cell densely populated bull semen microscopic observation video presents challenges that are more difficult than those presented by other general object-detection cases. These include partial occlusion, vast number of objects single frame, tiny size the object, artifacts, low contrast, resolution, and blurry because rapid movement sperm cells. This study proposes deep neural network architecture, called DeepSperm, solves aforementioned problems accurate faster state-of-the-art architectures. In proposed we use only one layer, which specific for small object detection. For handling overfitting increasing accuracy, set higher input dropout perform data augmentation on saturation exposure. Several hyper-parameters tuned to achieve better performance. Mean average precision (mAP), confusion matrix, precision, recall, F1-score used measure accuracy. Frame per second (fps) speed. We compare our method with you look once (YOLO) v3 YOLOv4. experiment, 94.11 mAP test dataset, 0.93, processing speed 51.9 fps. comparison YOLOv4, 2.18 x testing, 2.9 training while achieving comparative The weights file was also reduced significantly, one-twentieth Moreover, it requires 1.07 less graphical unit (GPU) memory simple, effective, efficient architecture its configuration detect cells robustly real time. experiments, surpass terms speed, resource needs.