A LOW-COST AUTOMATED ROTIFER Brachionus spp. COUNTING METHOD USING BACKGROUND SUBTRACTION AND CONNECTED COMPONENTS ANALYSIS

Rotifers Brachionus spp. are crucial starting diet for fish aquaculture. In commercial scale rotifer production, sampling and counting are essential for examining the rotifer population. The counting of rotifer samples is usually manually executed using microscopy by aquaculturists and is extremely time consuming and inefficient. Automation of rotifer counting would improve the production efficiency and release aquaculturists from tedious work. Current automated cell counting systems are typically very expensive and not designed for commercial scale rotifer culture. The present study developed a low cost computer vision system that can detect and count rotifers by analyzing microscopic images of rotifer samples using fine-tuned background subtraction (BS) and connected components labeling (CCL) algorithms. This method can accurately (error rate 5.16%) count rotifer samples (0.1 ml) from commercial scale production.

The image processing and analysis program was implemented using Python 3.7.3 and OpenCV 4.1.0. The testing procedure for this method is: 1) Take a 30s microscopic video clip of the 0.1ml rotifer sample. 2) Process the video frames with Gaussian Mixture Model based BS to generate foreground binary images. 3) Process the binary images with Scan plus Array-based Union-Find CCL to detect binary large object (blob). 4) Threshold the blobs by size to identify and count rotifers.

The equipment for data collection includes: iPhone7, microscope cell phone adapter (iDu LabCam, 10x), Leica GZ6 (1.1x). There are a couple of parameters such as the frame extraction stride,  background model learning rate, blob size threshold, etc. that were fine-tuned by the authors. The counting method was tested on 42 rotifer culture samples taken from University of Miami Experimental Hatchery (UMEH) with densities that ranged from 50 ind. ml^-1 to 450 ind. ml^-1. For each sample, the algorithm took and counted around 40 frames from each video clip and used the mean as the final count after removing the outliers. The method is also able to provide real-time counts for each frame of the video.

Figure 1 shows the performance of this method by comparing the counts using this method and manual counts of the Lugol-fixed samples. The performance index is the average absolute error rate (%) using the manual count as the ground truth. The R2 for fitting the data on y=x is 0.9934. The total cost of this method is only around $200 USD for the adapter assuming that most commercial hatcheries have the other tools such   laptop computers and microscopes.

The program can be obtained from the author by request via email