Hybrid CNN-GNN Models in Active Sonar Imagery: an Experimental Evaluation
Gabriel Arruda Evangelista, João Baptista de Oliveira e Souza Filho · 2024
The development of sonar technologies, such as Multibeam Forward Looking Sonar (MFLS), has enabled detailed underwater imaging, which can be applied for tasks like identifying mine-like objects. However, obtaining large datasets to train image recognition models remains challenging, leading to the need for smaller yet equally accurate alternative models. Previous research proposed a hybrid model that combines Convolutional Neural Networks with Graph Neural Networks for MFLS image classification. This study refines the feature extractor of this model using Knowledge Distillation (KD) and evaluates the cost-effectiveness of this pipeline compared to alternative solutions. The proposed method achieved an error rate of 6.42%, a value comparable to that of other solutions but with less computational effort.