Automated Recognition of Conidia of Nematode-Trapping Fungi Based on Improved YOLOv8

Abstract

Accurate identification of fungal species is crucial for mycological research, relying significantly on experienced taxonomists’ ability to recognize fungal morphology. With the dwindling number of taxonomists, developing a rapid, accurate, and automated fungal identification method is essential, promising to markedly decrease the time and resources needed for such tasks. To address the challenges in the automatic identification of nematode-trapping fungi (NTF, belong to Orbiliaceae, Orbiliomycetes, a group of filamentous fungi that catch nematodes in soil), specifically focusing on their conidia which are often aggregated and morphologically similar, we developed an enhanced method. Our approach integrates an efficient channel attention (ECA) mechanism within the backbone layer of YOLOv8l, improving the algorithm’s capability to detect global information effectively. This adjustment significantly optimizes the detection and differentiation of limited features and small-sized conidia among aggregated samples. The experimental outcomes demonstrated that our method increases precision by 0.3%, recall by 1.8%, and mean average precision (mAP) by 0.4%, surpassing the performance of existing target detection algorithms. In addition, this method accurately identified the aggregated fungal conidia and distinguished morphologically similar spores. In conclusion, the system accurately detects and recognizes the conidia of NTF in mixed spore environments, thus providing crucial technical support for the automatic detection and early prediction of NTF.

Keywords

• Conidia
• deep learning
• ECA attention mechanism
• nematode-trapping fungi
• YOLOv8