A lightweight few-shot marine object detection network for unmanned surface vehicles
Unmanned surface vehicles (USVs) are playing an important role in marine research, exploration and development. However, the perception capability of USVs is limited due to some objective factors. Firstly, the perception modules usually need to be deployed on processing devices with limited computin...
Gespeichert in:
Veröffentlicht in: | Ocean engineering 2023-06, Vol.277, p.114329, Article 114329 |
---|---|
Hauptverfasser: | , , , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Unmanned surface vehicles (USVs) are playing an important role in marine research, exploration and development. However, the perception capability of USVs is limited due to some objective factors. Firstly, the perception modules usually need to be deployed on processing devices with limited computing power, considering the power consumption and thermal dissipation. Secondly, collecting abundant training datasets of marine objects is a rather difficult mission due to complex environmental factors, such as illumination, weather and sea conditions. Thirdly, the perception modules cannot adapt to detection task of new objects fast enough, especially in few-shot scenarios they cannot perform well. To solve the above problems, we improved the ShuffleNet and designed Context Attention Enhancement FPN (CAE-FPN) to get an efficient lightweight network which is called ISDet. Then, a Progressive Gradient and Dynamic Learning Rate for MAML (PD-MAML) is proposed to solve the instability problem in meta training process with few-shot scenario. And a feature reweighting module is proposed to adapt our designed ISDet to new few-shot category. Experiments show that compared with other lightweight state-of-the-art networks, the proposed ISDet achieves better mean average precision with less model size, and it adapts to a new few-shot category fast while maintains the detection precision of existing categories.
•We design a lightweight marine object detection network ISDet, which achieves higher average precision with less parameters.•Ghost Block, SE Block, context enhancement module and attention mechanism are proposed to improve feature fusion efficiency.•We design a model agnostic meta learning method with progressive gradient for few-shot marine object detection task.•Feature reweighting module is proposed to adapt ISDet to few-shot categories and the performance is verified effectively. |
---|---|
ISSN: | 0029-8018 1873-5258 |
DOI: | 10.1016/j.oceaneng.2023.114329 |