Comparison of Convolutional Neural Network Architectures for Underwater Image Classification

Abstract

Convolutional neural networks (CNNs) play an essential role in classifying images collected in real-world environments. This article presents a performance comparison of selected CNNs for image classification tasks related to marine flora and fauna, using recordings from an Unmanned underwater vehicle (UUV). An attempt was made to find suitable CNN architectures for processing images of a poor-visibility marine environment among five commonly used architectures. The research was based on a uniform model training system: the same dataset and identical optimisation parameters were used to demonstrate the learning capabilities of each architecture. Thanks to the uniform CNN learning system, their direct learning capabilities for specific images can be more accurately estimated. This means that the conducted experiments showed that, in the early stage of training, the analysed networks achieved similar learning results, whereas the differences concerned the final training accuracy. The best results were achieved with models such as ResNet50, which have the most advanced architecture. Advanced models achieve improved classification of complex and distorted images by leveraging more parameters. The results provide insights into the performance of different architectures in underwater image classification and serve as a reference for further research on deep learning applications in marine environment monitoring.