sMarT RCI - Unsupervised Double K-Means Ensemble Clustering Tool for Remotely Operated Vehicle Camera Images

This work addresses the challenge of filtering irrelevant frames from Remotely Operated Vehicle (ROV) mission videos using advanced clustering techniques. We discuss the structure of a ROV surveillance mission. Furthermore we propose a method to differentiate between frames containing man-made structures and those without, employing a low-dimensional embedding generated by pre-trained neural networks. The clustering is performed using K-Means with a focus on two classes: “structure present” and “no structure present.” We evaluate preprocessing techniques, including Contrast Limited Adaptive Histogram Equalization (CLAHE) and grayscale conversion, to enhance clustering performance. The sMart RCI tool is introduced as an ensemble clustering approach that leverages multiple neural network backbones-InceptionV3, Vgg16,and VGG19-combined with grayscale and CLAHE preprocessing to improve overall accuracy. Our results demonstrate that ensemble clustering with double clustering approaches outperforms individual algorithms, providing consistent performance across different scenarios achiving an Fl Score of 92,7%. Comparative analysis with existing methods reveals significant improvements in precision and recall, particularly in varying conditions.