Classification of Underwater Fish Species Using Custom-Built Deep Learning Architectures

Fish are an integral part of the marine ecosystem. It has been noted that there are around 3.5 trillion; however, there are only 34,000 recognized species as of now. Classifying such a vast number of species manually from images can be very difficult due to the existence of noise, occlusion, distortion and an absence of illumination in the collected images. Advanced technological tools such as deep learning algorithms are being implemented to solve the challenge of image classification. The objective is to classify various species of fish by building deep learning models using various custom-built CNN architectures (VGG16, MobileNetV2, ResNet50) on the Fish4Knowledge data set. The Fish4Knowledge data set contains around 27 k images of fish belonging to 23 different classes. Each model was compared based on accuracy metrics and number of correctly classified classes. Based on the results, it was observed that MobileNetV2 performed the best, exhibiting an accuracy of 94.96% followed by VGG16 and ResNet50 at 91.61% and 85.71%, respectively. Hence, MobileNetV2 is the best that could be used for building a model to classify various species of fish and in turn help marine scientists. This can be explained by the fact that MobileNetV2 has a smaller number of parameters, which makes it much faster and more accurate in classifying fish.