A COMPARATIVE STRATIFICATION OF FISH SPECIES USING TRANSFER LEARNING ON PRE-TRAINED DEEP LEARNING NETWORKS JUXTAPOSED WITH SHUFFLERES – A HYBRID DEEP NETWORK CLASSIFIER

The marine ecoculture is an evolving realm that necessitates thorough scrutiny of the diverse species it
comprises, along with the explicit identification of the species classes that form, to be crucial for
aquaculture and the ecological conservation of fish diversity. The stratification through image
classification is a well-studied area of research using various conventional algorithmic methods.
However, the need to progressively identify deep features to stratify the multiple species of fish
unambiguously remains the pivotal study of this investigation. Deep learning methodologies utilize
various pre-trained networks that enable the identification of fish species through a systematic, layered
approach of non-linear activation functions, which delineate feature patterns and thereby achieve higher
classification accuracy. The study proposes a new method for stratifying fish species by applying transfer
learning to pre-trained deep learning networks, namely AlexNet, InceptionV3, and Resnet-18, along with
the application of ShuffleRes, a hybrid deep network classifier. To address the specified research
question, the work employs transfer learning, which enables the exploitation of knowledge from large
image datasets by fine-tuning pre-trained models. This approach enhances classification performance in
fish species, despite the limited availability of annotated data. Furthermore, the proposed ShuffleRes
architecture combines the advantages of residual connections and shuffle layers, promoting improved
feature representation, discriminative capacity, and surpassing classification accuracy compared to
individual pre-trained networks. The simulations are implemented in MATLAB, and the results for the
study are successfully procured.