Semi-automatic one-class image labeling using a neural network object detection model

As a result of a sharp popularity increase of artificial intelligence resource-intensive methods, a serious problem arises in the preliminary data preparation for the convolutional neural networks models effective training. The authors present an approach based on the training dataset iterative updating principle using the YOLO neural network model for areas of interest detection, objects selection, and the original images labeling process automation. The proposed approach was tested with various model configurations for bacteria labeling on images obtained using a scanning electron microscope and, on average, demonstrated ~90% precision on a training dataset increased by 1.75 times over the initial training dataset.