Classification of Fish Species with Augmented Data using Deep Convolutional Neural Network

Abstract

Classification is one of the primary data science tasks involving large datasets. To date, fish species classification in the Philippines is considered significant for further aquaculture and protection. However, immense efforts and knowledge are necessary to determine fish characteristics through classification. The VGG16 network is one of the top pre-trained models but is still not able to accurately classify common fish species found in Verde Island. This study primarily aims to classify Verde Island fish species using a modified VGG16 network. The VGG16 Deep Convolutional Neural Network (DCNN) undergoes retraining, fine-tuning, and optimization to provide better accuracies in classifying specific Verde Island fish species. Also, this research generated augmented synthetic data for training and testing the model, as there are limited images available. Augmented images are flipped, rotated, cropped, zoomed, and sheared to provide a robust number of features for classification. Results of the training the model achieves 99 percent accuracy for the three different fish species. Hence, this study concludes that a pre-trained model like VGG16 can still improve by fine-tuning, optimization, and data augmentation to classify specific fish species. This paper also includes possible future works determined by the authors.