Microalgae Detection Based on Cascade R-CNN Object Detection Model

Abstract

Marine microalgae are one of the significant biological resources in marine ecosystems and a part of the “blue carbon sink.” Artificial identification of marine microalgae usually takes a lot of time, so using the object detection method to detect microalgae automatically can save a lot of artificial resources. The official website provides an algae dataset in the IEEE UV 2022 “Vision Meets Algae” object detection challenge. However, this dataset contains many small objects, which is unfavorable for the object detection model to identify algae. We use Cascade R-CNN with the backbone ConvNeXt-B as our main object detection model in this challenge. To make the model recognize small objects well, we increase the input image size and add global context to the model. During training, we used data augmentation and multi-scale training strategies that improved the performance of the model. Finally, to improve the detection performance, we integrate Cascade R-CNN, TOOD, and GFL. We evaluated our method on the test set. The mAP of Cascade R-CNN reached 54.69, while the mAP of model integration reached 56.22.