Underwater object detection method based on learnable query recall mechanism and lightweight adapter

Xi Lin ¹, Xixia Huang ¹, Le Wang ¹

⁰Institute of Logistics Science and Engineering, Shanghai Maritime University, Shanghai, People's Republic of China.

Plos One +

|

February 28, 2024

View abstract on PubMed

Summary

This study enhances underwater object detection by improving DETR with a query recall mechanism and multi-scale features, significantly boosting performance in challenging marine environments.

Area Of Science

Marine science
Computer vision
Oceanography

Background

Underwater object detection is crucial for aquaculture, environmental monitoring, and marine science.
Existing deep learning models struggle with underwater image challenges like noise, blur, and multi-scale objects.

Purpose Of The Study

To adapt the DETR (DEtection TRansformer) model for improved underwater object detection.
To address limitations in current algorithms for detecting small, irregular, and noisy underwater targets.

Main Methods

Introduced a learnable query recall mechanism to reduce noise impact.
Developed a lightweight adapter for multi-scale feature extraction in encoding and decoding.
Optimized bounding box regression using a combination of smooth L1 and CIoU loss.

Main Results

The proposed method demonstrated significant improvements in underwater object detection.
Validated effectiveness against state-of-the-art methods on the RUOD dataset.
Achieved superior performance in handling noise, blur, and multi-scale objects.

Conclusions

The enhanced DETR model effectively addresses challenges in underwater object detection.
The proposed modifications offer a robust solution for marine visual tasks.
This work advances the capabilities of deep learning in underwater imaging applications.