Computer vision system to count crustacean larvae

TL;DR

This study developed and tested two computer vision systems using different cameras and illumination conditions to automatically count crustacean larvae in industrial ponds, achieving over 86% accuracy.

Contribution

The paper introduces a novel application of YOLOv5 CNN for crustacean larvae counting and models larvae growth using the Gompertz function.

Findings

Achieved 88.4% detection accuracy with iPhone 11 system

Achieved 86% detection accuracy with DSLR camera system

Developed a larvae growth model with R squared of 0.983

Abstract

Fish products account for about 16 percent of the human diet worldwide, as of 2017. The counting action is a significant component in growing and producing these products. Growers must count the fish accurately, to do so technological solutions are needed. Two computer vision systems to automatically count crustacean larvae grown in industrial ponds were developed. The first system included an iPhone 11 camera with 3024X4032 resolution which acquired images from an industrial pond in indoor conditions. Two experiments were performed with this system, the first one included 200 images acquired in one day on growth stages 9,10 with an iPhone 11 camera on specific illumination condition. In the second experiment, a larvae industrial pond was photographed for 11 days with two devices an iPhone 11 and a SONY DSCHX90V cameras. With the first device (iPhone 11) two illumination conditions were tested. In each condition, 110 images were acquired. That system resulted in an accuracy of 88.4 percent image detection. The second system included a DSLR Nikon D510 camera with a 2000X2000 resolution with which seven experiments were performed outside the industrial pond. Images were acquired on day 1 of larvae growing stage resulting in the acquisition of a total of 700 images. That system resulted in an accuracy of 86 percent for a density of 50. An algorithm that automatically counts the number of larvae was developed for both cases based on the YOLOv5 CNN model. In addition, in this study, a larvae growth function was developed. Daily, several larvae were taken manually from the industrial pond and analyzed under a microscope. Once the growth stage was determined, images of the larva were acquired. Each larva's length was measured manually from the images. The most suitable model was the Gompertz model with a goodness of fit index of R squared of 0.983.