Image Filtering Model Based on Convolutional Neural Networks for Automatic Counting of Post-larvae in Aquaculture

Aquaculture is an important activity for the animal protein global supply. In Brazil, this area is showing significant growth in recent years. Among the activities carried out during the production process, the counting of animals in the initial stages can be highlighted. A large number of Brazilian aquaculture farms are small, turning difficult to acquire novel solutions for the automatic count of post-larvae. To mitigate this issue, this paper intends to develop a solution based on the performance evaluation of a set of counting models for use in embedded structures. In addition, this application can be scalable for counting different species and sizes. Besides, the dataset named Vivarium and its specifications is presented as proof of concept and used in the model evaluation. Results show that the prediction model based on convolutional neural networks is capable of verifying the compliance of images, achieving an accuracy of 99%.