Microscopic Studies of Activated Sludge Supported by Automatic Image Analysis Based on Deep Learning Neural Networks
Więcej
Ukryj
1
Department of Applied Mathematics, Faculty of Mathematics and Information Technology, Lublin University of Technology, Nadbystrzycka 38, 20-618 Lublin, Poland
2
Department of Water Supply and Wastewater Disposal, Faculty of Environmental Engineering, Lublin University of Technology, Nadbystrzycka 40B, 20-618 Lublin, Poland
Autor do korespondencji
Grzegorz Łagód
Department of Water Supply and Wastewater Disposal, Faculty of Environmental Engineering, Lublin University of Technology, Nadbystrzycka 40B, 20-618 Lublin, Poland
J. Ecol. Eng. 2024; 25(4):360-369
SŁOWA KLUCZOWE
DZIEDZINY
STRESZCZENIE
Paper presents a microscopic studies of activated sludge supported by automatic image analysis based on deep learning neural networks. The organisms classified as Arcella vulgaris were chosen for the research. They frequently occur in the waters containing organic substances as well as WWTPs employing the activated sludge method. Usually, they can be clearly seen and counted using a standard optical microscope, as a result of their distinctive appearance, numerous population and passive behavior. Thus, these organisms constitute a viable object for detection task. Paper refers to the comparison of performance of deep learning networks namely YOLOv4 and YOLOv8, which conduct automatic image analysis of the afore-mentioned organisms. YOLO constitutes a one-stage object detection model that look at the analyzed image once and allow real-time detection without a marked accuracy loss. The training of the applied YOLO models was carried out using sample microscopic images of activated sludge. The relevant training data set was created by manually labeling the digital images of organisms, followed by calculation and comparison of various metrics, including recall, precision, and accuracy. The architecture of the networks built for the detection task was general, which means that the structure of the layers and filters was not affected by the purpose of using the models. Accounting mentioned universal construction of the models, the results of the accuracy and quality of the classification can be considered as very good. This means that the general architecture of the YOLO networks can also be used for specific tasks such as identification of shell amoebas in activated sludge.