Computer vision in aquaculture: a case study of juvenile fish counting

In aquaculture breeding or production programmes, counting juvenile fish represents a considerable cost in terms of the human hours needed. In this study, we explored the use of two state-of-the-art machine learning architectures (Single Shot Detection, hereafter SSD and Faster Regions with convolutional neural networks, hereafter Faster R-CNN) to augment a manual image-based juvenile fish counting method for the Australasian snapper (Chrysophrys auratus) bred at The New Zealand Institute for Plant and Food Research Limited. We tested model accuracy after tuning for confidence thresholds and non-maximal suppression overlap parameters, and implementing a bias correction using a Poisson regression model. Validation of image data showed that after tuning, bias-corrected SSD and Faster R-CNN models had mean absolute percent errors (MAPE) of less than 10%, with SSD having MAPE of less than 5%. Comparison of the results with those from manual counts showed that, while manual counts are slightly more accurate (MAPE = 1.56), the machine learning methods allow for more rapid assessment of counts and thus facilitating a higher throughput. This work represents a first step for deploying machine learning applications to an existing real-life aquaculture scenario and provides a useful starting point for further developments, such as real-time counting of fish or collecting additional phenotypic data from the source images.

Automated summary

This study explored the use of machine learning architectures to improve the accuracy and efficiency of juvenile fish counting in aquaculture breeding. The results showed that the machine learning models had high accuracy after tuning, enabling rapid assessment of fish counts and facilitating higher throughput. This work provides a useful starting point for further developments in real-time counting and data collection.