Performance Comparison of Convolutional Neural Network-based model using Gradient Descent Optimization algorithms for the Classification of Low Quality Underwater Images

Underwater imagery and analysis plays a major role in fisheries management and fisheries science helping developing efficient and automated tools for cumbersome tasks such as fish species identification, stock assessment and abundance estimation. Majority of the existing tools for analysis still leverage conventional statistical algorithms and handcrafted image processing techniques which demand human interventions and are inefficient and prone to human errors. Computer vision based automated algorithms need a better generalisation capability and should be made efficient to address the ambiguities present in the underwater scenarios, and can be achieved through learning based algorithms based on artificial neural networks. This paper research about utilising the Convolutional Neural Network (CNN) based models for under water image classification for fish species identification. This paper also analyses and evaluates the performance of the proposed CNN models with different optimizers such as the Stochastic Gradient Descent (SGD),Adagrad, RMSprop, Adadelta, Adam and Nadam on classifying ten classes of images from the Fish4Knowledge(F4K) database.