Identification and counting of pirapitinga Piaractus brachypomus fingerlings fish using Machine learning with batch normalization

Abstract

Abstract The identification and counting of fish are relevant tools used for stocking, harvesting, and marketing management of farmed fish. The use of convolutional networks has been used for such purposes, and different approaches have been employed to improve network learning. Batch normalization is one of the techniques that aids in the enhancement and stability of the network, providing greater accuracy. Thus, the objective was to evaluate machine learning in the identification and counting of pirapitinga Piaractus brachypomus fingerlings fish with and without batch normalization. One thousand photographic images of pirapitinga fingerlings were used, labeled through bounding boxes. The training of the adapted convolutional network model was performed, with batch normalization layers inserted at the end of each convolution block. One hundred fifty epochs were established, and batch sizes for normalization were set to 1, 5, 10, and 20. The database training was also conducted without applying normalization for comparison. The evaluation metrics for network performance were precision, recall, and mAP@0.5. The results obtained with the model without the application of the technique were inferior to the models in which batch normalization was applied. The batch size equal to 20 was the model trained with the best performance, showing precision of 96.74%, recall of 95.48%, mAP@0.5 of 97.08%, and accuracy of 98%. It is concluded that batch normalization increases accuracy in the detection and counting of pirapitinga fingerlings in different densities of fish