Dynamically Tunable Multidimensional Feature Focusing and Diffusion Networks for Water Surface Debris Detection
With the increasing level of industrialization, water surface debris detection has emerged as a significant research topic. However, challenges such as complex backgrounds, overlapping targets, and reflections on water surfaces often hinder effective waste detection, resulting in low model accuracy...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11122446/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the increasing level of industrialization, water surface debris detection has emerged as a significant research topic. However, challenges such as complex backgrounds, overlapping targets, and reflections on water surfaces often hinder effective waste detection, resulting in low model accuracy and unreliable recognition. To address these challenges, the Water Surface Debris DEtection TRansformer (WSD-DETR) was proposed. First, a Self-moving Point Convolutional Gating Network (SPCG-Net) was designed, which integrated an adaptive point-moving mechanism with a convolutional gating linear unit to enhance the flexibility and accuracy of feature extraction. Second, an Attention-based Re-parameterization of Intra-scale Feature Interactions (ARIFI) module was constructed to process high-level features extracted from the backbone. This module employed a single-scale transformer encoder with Re-parameterized Batch Normalization (RepBN) to improve focus on small and medium-sized targets in water surface waste detection, thereby capturing relationships between semantic concepts and conceptual entities. Furthermore, a Focal-Diffuse Feature Pyramid Network (FD-FPN) was introduced to accurately capture and integrate key feature information through focused feature fusion techniques while utilizing cross-scale diffusion analysis to efficiently transfer and enhance feature information across different scales. This approach significantly improved feature expression capability and overall model performance. Experimental results indicated that WSD-DETR achieved a precision of 88.1%, and reduced model parameters by 12.6% compared to the Real-Time DEtection TRansformer (RT-DETR), and increased mAP@50 and mAP@50:90 values by 4.4% and 5.5%, respectively. These outcomes demonstrated substantial potential for water surface debris detection. |
|---|---|
| ISSN: | 2169-3536 |