Automatic ship detection in maritime environments remains a challenging task in computer vision, mainly due to the variability of environmental conditions, water surface reflections, and the diversity of vessel profiles and proportions. This dissertation proposes the development of a lightweight and efficient system for ship detection in optical images, named SOD-Edge (Single Object Detection for Edge Computing), targeting deployment in embedded systems with limited energy resources. The methodology combines feature extraction from efficient convolutional neural networks with Weightless Neural Networks (WiSARD) for classification and multiple regression strategies for bounding box estimation, including Support Vector Regression (SVR), k-Nearest Neighbors (kNN), and a regression-based variant of WiSARD (ReWiSARD).
Experiments were conducted in two different computational environments: a cloud-based setup (Google Colab) and an embedded device (Raspberry Pi 4). The evaluation considered not only performance metrics such as accuracy, F1-score, MAE, and IoU, but also efficiency regarding runtime and energy consumption. Results show that the combination of efficient CNNs, WiSARD, and lightweight regression methods provides competitive performance compared to purely deep learning-based approaches, while offering advantages in terms of structural simplicity, runtime efficiency, and reduced energy consumption.
This study highlights the relevance of exploring alternatives to conventional deep neural networks in constrained computational scenarios and demonstrates that the choice of the regression method can be guided by application-specific requirements. Finally, it opens perspectives for future extensions involving multi-object detection and video analysis in maritime environments.