Ocean Fish Classification using pre-train CNN models and Performance Evaluation

Ocean fish are essential for marine biodiversity and support ecosystem health and sustainable fisheries and local economies. Properly classifying the fish is necessary to keep track of the fish population and encourage sustainable fisheries management. This study attempts to tackle the arduous task of fish species classification through eight pre-trained Convolutional Neural Network (CNN) models, using the Croatian Fish Dataset with 794 images divided into 12 classes. Due to the small size and class imbalance in the data set, preprocessing techniques such as resizing, normalization and denoising as well as advanced data augmentation techniques were applied to improve model performance. ResNet50 was the best performing model with 89.48% accuracy due to the fact that it has a strong architecture built with residual connections that help with gradient vanishing as well as allowing it to learn deeper. DenseNet121 ranked right after with an accuracy of 83.86%, which offers a dense connectivity that facilitates feature reuse and gradient flow. This study sets important standards for using small, uneven datasets in ecological research, showing how deep learning can help protect biodiversity. The results highlight the need to choose models that fit the unique challenges of the data. It also points to future research paths, like improving data augmentation, adjusting hyperparameters, and using ensemble methods to boost classification accuracy and make a bigger difference in ecology.