Maritime Small Object Detection Algorithm in Drone Aerial Images Based on Improved YOLOv8
Combining unmanned aerial vehicles (UAVs) with deep learning algorithms offers an efficient, safe and inexpensive alternative to maritime search and rescue (mSAR) missions. Maritime UAV images present unique challenges for object detection due to their complex nature, including dense distribution, m...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10741529/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Combining unmanned aerial vehicles (UAVs) with deep learning algorithms offers an efficient, safe and inexpensive alternative to maritime search and rescue (mSAR) missions. Maritime UAV images present unique challenges for object detection due to their complex nature, including dense distribution, multi-scale objects and occlusion. Aiming to address this problem, we propose a novel lightweight model specifically designed for maritime small object detection, named AB2D-YOLO. Firstly, the attention based intra-scale feature interaction (AIFI) module is used to replace the spatial pyramid pooling fast (SPPF) module on the backbone, enhancing the detection precision of occluded and densely small targets by integrating global and contextual feature information. Secondly, the dilation-wise residual (DWR) module is integrated into the network. The module employs three sets of dilated convolution with different sampling rates to obtain multi-scale receptive fields, which effectively improves the capacity for detecting multi-scale objects. Then, we propose an improved network fusion model based on weighted bi-directional feature pyramid network (BiFPN) to reconstruct the neck, which can enhance the features of small targets through weighted fusion of feature information of different scales and bidirectional cross-scale connection. Finally, we add a new detection layer in the neck to capture more object location information in images. When compared to the benchmark model YOLOv8s, AB2D-YOLO achieves an 8.96% increase in mean average precision (mAP) on the SeaDroneSee dataset, while maintaining a low model complexity with only 6.95 MB of parameters. When compared to state-of-the-art models, AB2D-YOLO model is conducive to the deployment of maritime UAV. |
|---|---|
| ISSN: | 2169-3536 |