Automatic counting of fish larvae using computer vision based on neural networks

Abstract

Fish larvae counting is a technique applied in aquaculture to determine the efficiency of the induction stage and to know the number of fertilized larvae. For this reason, the research aims to improve the count of larvae under 3 fundamental pillars: precision, error and time. For this, we carried out an experimental investigation under a completely randomized design with two counting systems: traditional and artificial vision; each one with 10 repetitions, with 2000 larvae; Later, we carried out the count by means of artificial vision using a camera that captured images of a fish tank with moving fish. The results show that the proposed method is feasible for counting larvae, with 92.65% accuracy, 7.41% error and an average time of 61 seconds per repetition in relation to the traditional counting system: accuracy 64.44%, error 35.61% and time 2009.3 s. The developed system can be replicated in the aquaculture sector due to its effectiveness and cost.