YOLOx model-based object detection for microalgal bioprocess

Microalgae can fix carbon dioxide from flue gases and utilize nutrients from wastewater while producing valuable biomass production of biofuels, chemicals, and fertilizers. Microalgae detection is of great significance for identifying mixed algae species in nature. Unfortunately, current manual detection methods are timeconsuming, low in precision, and poor in universality. A YOLOx-s-based microalgae detection method was first proposed in this study, which uses a multi-scale and multi-morphology microalgae (Chlorella, Scenedesmus, and Spirulina species) image dataset as model input. When Focal Loss was selected for the classification loss, the number imbalance problem in microalgal detection was solved. As DIoU Loss was employed in regression loss, tiny-scale microalgae were well detected with shortened 20.24 % processing time, and the ASFF module benefited the overall model performance by fusing features. The Precision and Recall of improved YOLOx-s in detecting microalgae achieved 95.93 % and 93.48 %, and the mAP was improved by 3.33 % compared to the original YOLOx-s.

Automated summary

A YOLOx-s-based microalgae detection method was proposed in this study, which utilized a multi-scale and multi-morphology microalgae image dataset as model input. The classification imbalance problem in microalgal detection was solved by selecting Focal Loss for the classification loss. By employing DIoU Loss in regression loss, tiny-scale microalgae were well detected with a shortened processing time. The overall model performance was improved by the ASFF module through feature fusion, and the Precision and Recall of the improved YOLOx-s in detecting microalgae achieved 95.93% and 93.48%, respectively, with a 3.33% improvement in mAP compared to the original YOLOx-s.