PR-EWC: Period Replay and EWC for Class-Incremental Maritime Vessel Detection

View at publisher Summarise with VIRAT
Authors: Hongyan Guo, Junjian Feng, Bing Zheng, Chao Dong
Cited by: 0
References: 19

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Traditional object detection approaches tend to perform poorly in dynamically changing environments, as they are typically trained on static datasets that fail to capture real-world variability. However, these methods are particularly prone to catastrophic forgetting—the phenomenon in which learning new object categories significantly degrades performance on previously acquired knowledge. To address this issue, we propose PR-EWC, an incremental learning-based approach for maritime vessel detection. At the parameter level, we incorporate an Elastic Weight Consolidation (EWC) mechanism, which quantifies the importance of each parameter using the Fisher Information Matrix and applies parameter-specific regularization to mitigate catastrophic forgetting during incremental learning. At the data level, we adopt a data replay strategy, leveraging a periodic replay mechanism to reinforce both newly learned and previously acquired knowledge. Experimental results on the SeaShips dataset demonstrate that the proposed method effectively mitigates catastrophic forgetting. Compared with direct incremental learning using YOLO11n, which achieves a mAP50 of only 23.4%, the proposed approach significantly improves performance, reaching a mAP50 of 79.90% and showing a remarkable enhancement in performance.

Fields of Study & Subjects

Domain Adaptation and Few-Shot Learning Advanced Neural Network Applications Data Stream Mining Techniques

Publication details

Journal / source
International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
Pages
1024–1029
Publication date
16 January 2026
Language
EN