Navigating Coreset Selection and Model Compression for Efficient Maritime Image Classification

Efficient classification of maritime images is crucial for applications in surveillance, environmental monitoring, and navigation, especially in resource-constrained environments such as edge devices and low-power systems. While deep neural networks (DNNs) perform well on datasets like the Maritime Satellite Imagery (MASATI) dataset, their high computational and memory requirements hinder deployment on constrained hardware. Additionally, maritime environments present challenges such as dynamic backgrounds, occlusions, and variable weather conditions, complicating the classification task. Although, coreset selection techniques reduce training costs by focusing on key samples, and model compression mitigates deployment challenges, yet their combined potential remains untapped. This work proposes a novel framework that integrates coreset selection and model compression to optimize maritime image classification for resource-limited settings. By employing coreset selection techniques (random, margin-based, and forgetting-based) in conjunction with post-training pruning and mixed-precision quantization, the framework reduces model size by up to 94% through quantization and lowers computational complexity by 50% via coreset selection, all while maintaining approximately 95% classification accuracy. These results demonstrate the framework's robustness in diverse maritime conditions and highlight its potential for efficient, scalable maritime image classification on edge and mobile devices.