Underwater Moving Object Detection using an End-to-End Encoder-Decoder Architecture and GraphSage with Aggregator and Refactoring

Abstract

Underwater environments are greatly affected by several factors, including low visibility, high turbidity, backscattering, dynamic background, etc., and hence pose challenges in object detection. Several algorithms consider convolutional neural networks to extract deep features and then object detection using the same. However, the dependency on the kernel’s size and the network’s depth results in fading relationships of latent space features and also are unable to characterize the spatial-contextual bonding of the pixels. Hence, they are unable to procure satisfactory results in complex underwater scenarios. To re-establish this relationship, we propose a unique architecture for underwater object detection where U-Net architecture is considered with the ResNet-50 backbone. Further, the latent space features from the encoder are fed to the decoder through a GraphSage model. GraphSage-based model is explored to reweight the node relationship in non-euclidean space using different aggregator functions and hence characterize the spatio-contextual bonding among the pixels. Further, we explored the dependency on different aggregator functions: mean, max, and LSTM, to evaluate the model’s performance. We evaluated the proposed model on two underwater benchmark databases: F4Knowledge and underwater change detection. The performance of the proposed model is evaluated against eleven state-of-the-art techniques in terms of both visual and quantitative evaluation measures.