Challenges in Underwater Video Data Processing

This study presents a detailed pipeline for improving and analyzing images using several computational techniques, such as Exploratory Data Analysis (EDA), Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Gradient-Weighted Class Activation Mapping (Grad-CAM) visualizations. The purpose of this research is to standard-ize underwater video datasets, spot important features, and im-prove the quality of noisy images. Pre-processing was performed on the dataset by resizing and denoising to make sure that it is suitable for future tasks by making sure that it is uniform and compatible for processing. EDA provided valuable information on the variability of the datasets, while the dimensionality reduction methods of PCA and t-SNE showed that certain patterns or clusters of data points naturally form in the low-dimensional space, indicating underlying relationships or similarities within the dataset. Using DenseNet121 and Grad-CAM, features were extracted and carefully visualized to learn how the DenseNet121 model makes its predictions. The techniques in our study can be used to further enhance the quality of underwater images for environmental/ecological health and biodiversity monitoring.