Object Detection and Abundance Analysis for Fountain-Flow Imaging of Marine Plankton

The special optofluidic configuration of a fountain-flow imaging system is primarily aimed to reduce motion blur associated with side-view flow imaging at high speed, facilitating subsequent analysis of the plankton sample taxonomy and abundance with better efficiency. However, repeated recording of the same objects and defocusing blur in an image sequence acquired by a darkfield fountain-flow imaging system can lead to plankton recognition difficult and abundance statistics inaccurate. To resolve this problem, we propose a pipeline of processing algorithms for object counting in this paper, which is based on three steps of image preprocessing, object detection, and object tracking. After the raw images are preprocessed, a modified Faster R-CNN with Feature Pyramid Network is applied to obtain the category and location of all the planktonic organisms. Then the same object organism is tracked through consecutive frames based on their color and structural similarity, so that associated data can be eventually merged for counting repetition correction. The proposed counting method is tested on cultured and natural water samples containing different plankton species in laboratory and field. The test results have demonstrated that the detection accuracy and tracking accuracy of the method reach to 73% and 94%, respectively. It is expected the classification and counting of marine plankton from fountain-flow imaging of high density water samples could be analyzed more quickly and accurately by using the proposed method, which can further improve the observation capability for scientific research and routine monitoring of plankton.