Development of an image-based identification systems for freshwater diatoms using computer vision techniques
2024
Geronimo, J.O.V.
Diatoms play a crucial role in monitoring environmental health. Their rapid response and high sensitivity make them valuable indicators of pollution and habitat restoration success. Traditional identification and classification of diatoms rely on molecular genetic techniques and morpho-taxonomic characterization, known for being laborious and time-consuming. To address this, an automated detection and classification system employing YOLO architecture and digital image processing was proposed. The study demonstrates the system's success in recognizing and categorizing 16 genera of freshwater diatoms, showcasing variations in color, size, and shape. The initial phase focused on detection and localization using YOLOv4, yielding a promising mean average precision score of 56.64%. Four different experimental setups for identifying 16 diatom genera showed that YOLOv4 outerperformed YOLOv7, achieving a mean average precision of 19.21%. The observed impact of imbalanced and limited diatom cells in certain classes during the 16-class setup prompted a refined subsequent experiment focusing on eight genera with the highest number of cells. YOLOv4, specifically the model trained on the raw image dataset, showed significant improvement, reaching 41.98%. Comparing the automated approach with traditional methods revealed a substantial time difference. The automated system identified and counted diatoms in approximately 100 images within seconds to minutes, contrasting with the traditional method requiring at least two days for a batch of around 50 images. This search signifies a significant advancement in efficient diatom analysis, paving the way of robust systems capable of identifying diverse diatom species in the Philippines.
显示更多 [+] 显示较少 [-]