Identify and segment microalgae in complex backgrounds with improved YOLO

Abstract

Microalgae are widely distributed in the ocean, and some species are prone to causing harmful algal blooms that threaten the marine ecological environment. At present, microscopy is the most common method for microalgae analysis, and the combination of computer vision and microscopy is the mainstream trend in algae morphology classification. However, most methods only focus on the species and quantities of algae, without obtaining contour information that can further analyze their survival status and biomass. Therefore, this article proposes a convolutional neural network (AlgaeSeg-YOLO), which can recognize the species, quantities, and contours of algae. Firstly, a dataset of microalgae microscopic images in microfluidic chips was constructed, which includes a total of 2799 annotated images of 6 harmful microalgae, including 3916 instances. Secondly, the AlgaeSeg-YOLO was constructed with stronger feature fusion and pixel-level spatial information modeling capabilities based on YOLOv8n-seg. Finally, compared to other common methods, the experimental results show that the mAP (mean Average Precision) of AlgaeSeg-YOLO reaches 95.61 %, which is 1.64 %, 1.76 %, 5.28 %, 3.34 %, and 5.39 % higher than YOLOv8n-seg, YOLOv5n-seg, Mask R-CNN, Cascade Mask R-CNN, and SOLOv2, respectively, achieving real-time and accurate segmentation of microalgae in complex backgrounds. Meanwhile, the parameters and computation remain relatively low. This work helps to achieve fully automated analysis of microalgae, reduce labor costs, and provide technical support for long-term monitoring of the ecological environment and further research.