YOLOv8-based neural network algorithm for detecting biological objects: Daphnia magna case study
A.S. Olkova, E.V. Medvedeva
Section: Methodology and research methods. Models and forecasts
Automation of routine biological analysis procedures is an important interdisciplinary issue. The article describes a neural network algorithm for detecting meso-sized biological objects. The biomodel was Daphnia magna, which is frequent for bioassaying of natural and man-made environments. The algorithm is implemented on YOLOv8s lightweight convolutional neural network (CNN). Daphnia magna’s original photo and video images as well as those publicly available on the Roboflow and Kaggle platforms were used for CNN training and testing. The image database consisted of 12540 images, of which 430 were original, the rest were transformed using the functions of the Roboflow service. The training was conducted over 150 epochs, with an image resolution of 1280×1280 pixels. Procedures for counting and tracking objects are implemented using the built-in functions “Object Tracking” and “Object Counting”. The algorithm detects objects in static images and videos in real time. The processing speed of video images was about 50 ms per frame, which is enough for the algorithm to work in real time. The values of the algorithm quality rating metrics were the following: mAP – 89.8%; precision – 88.4%; recall – 87.4%. The main mistake of the neural network was counting the reflections of the same daphnia on the inner and outer surfaces of the aquarium glass. The algorithm is aimed at reducing the complexity of biotesting methods and increasing the accuracy of data processing results. The proposed neural network algorithm for detecting biological objects can be adapted through similar training to the search and counting of other organisms.
ISSN 1995-4301
Select viewing options