Tuna classification using super learner ensemble of region-based CNN-grouped 2D-LBP models

Abstract

Tuna is superior among the marine fishes that are exported in the forms of raw fish and processed food. Separation of Tuna into their species is done in industries manually, and the process is tiresome. This work proposes an automated system for classifying Tuna species based on their images. An ensemble of region-based deep neural networks is used. A sub region contrast stretching operation is applied to enhance the images. Each fish image is then divided into three regions and is augmented before giving as input to pre-trained convolutional neural networks (CNN). After fine-tuning the models, the output from the last convolutional layer is given to a grouped 2D-local binary pattern descriptor (G2DLBP). Statistical features from the descriptor are applied to different classifiers, and the best classifier for each image region model is identified. Different ensemble methods are subsequently used to combine the three CNN-G2DLBP models. Among the ensemble techniques, super learner ensemble method with random forest (RF) classifier using 5-fold cross-validation shows the highest classification accuracy of 97.32%. The performance of different ensemble methods is analyzed in terms of accuracy, precision, recall, and f-score. The proposed system shows an accuracy of 93.91% when evaluated with an independent test dataset. An ensemble of region-based CNN with textural features from G2DLBP is applied for the first time for fish classification.

Keywords

• Tuna classification
• Convolutional neural network
• Grouped 2D-local binary pattern
• Super learner ensemble