An Image Classifier for Underwater Fish Detection using Classification Tree-Artificial Neural Network Hybrid

Fish detection using imaging technologies and computer vision systems is considered as an effective tool in fish monitoring for increasing the production to satisfy future global demands. This persistent tasks, with image classification as one of its subtasks, encounters challenges due to the complex nature of underwater images. A proposed approach to address this subtask was to create a hybrid image classification model from classification tree and artificial neural network. The classification tree component performed feature selection to extract a reduced representation of the fundamental dataset, derived from a series of acquired and processed underwater images in a land-based aquaculture setup. This said representation was then fed to a feedforward artificial neural network to develop such model. The best configuration of this hybrid model was determined, based on learning time and cross entropy, and was compared to a classification tree and an artificial neural network, both developed from the fundamental dataset, based on training and testing accuracies. The best performing hybrid model, composed of 100 hidden neurons in the artificial neural network component, achieved training and testing accuracies of 93.6% and 78.0%, respectively, hence, providing a competitive solution to the image classification in fish detection problem.