An ensemble of region-based CNN models combined by sum rule for tuna classification

Among the most widely traded fish in the world, Tuna is an important fish due to their nutritional value. In the export industry, these fish are manually sorted into various species. In this paper, an automated system is suggested to speed up these industries in order to satisfy the growing need for tuna and its products. Three species of tuna are used here. The contrast of each image is enhanced by using a preprocessing step. Head, abdomen and tail region images are then split from the images. Different parameters then augment these images to get a class balanced dataset. The resulting augmented images are applied to four pre-trained models. The sum rule combines the predictions from the four networks for a region. The combined scores of three region-based predictions are used in the majority voting ensemble to get the final image-wise prediction. Different performance measures are used to evaluate the system, and results show that the proposed system classifies tuna fishes with an accuracy of 95.65%.