Aims / Objectives: Identification of fish species is essential in export industries. Among the different fish species exported, tuna forms a significant portion and hence the separation of tuna from other fishes is necessary. The work aims to develop automated systems for the separation of commercially important tuna from other fishes.
Methodology: The work proposes two models for the classification of commercial fishes. The first model uses conventional feature descriptors, which extract features from both spatial and frequency domain. These features are combined and are reduced by an ensemble dimension reduction method. The combined and reduced feature sets are evaluated using different classifiers. The second proposed model uses four pre-trained convolutional neural networks, VGG16, VGG19, Xception, and MobileNet, for the classification. The models are fine-tuned for the classification process.
Results: Results show that for the first model, extreme learning machine classifier with Mercer wavelet kernel gives high accuracy on combined feature set while the polynomial kernel ELM provides better performance with the reduced set. For the second model, a comparison of the performance of four CNN models is done, and results indicate that VGG19 outperforms other networks in the classification task.
Conclusion: Among the two proposed models, pre-trained CNN based model shows better performance than the conventional method in the separation task. Different performance measures, accuracy, precision, recall, F-score, and misclassification error are used to evaluate the system. A comparison of performance of the proposed models with the state-of-the-art systems is also reported.
Automated Classification of Glandular Tissue by Statistical Proximity Sampling
