Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Underwater object detection in sonar imagery with detection transformer and Zero-shot neural architecture search

XiaoTong Gu, Shengyu Tang, Yiming Cao, Changdong Yu

TL;DR

This work tackles underwater sonar object detection, where low resolution and high noise hinder performance. It introduces NAS-DETR, an entropy-guided zero-shot NAS framework that designs a CNN–Transformer backbone and couples it with a Deformable DETR–style detector, augmented by a content–position decoupled query initialization and a multi-task loss. The approach provides state-of-the-art results on URPC2021 and URPC2022 while preserving real-time speed, and it offers interpretability through Spearman-based analysis linking backbone depth and width to differential entropy. TensorRT quantization further boosts inference speed, making NAS-DETR practical for industrial deployment, with future work focusing on lightweight search, multi-modal fusion, and robust adaptation to dynamic environments.

Abstract

Underwater object detection using sonar imagery has become a critical and rapidly evolving research domain within marine technology. However, sonar images are characterized by lower resolution and sparser features compared to optical images, which seriously degrades the performance of object detection.To address these challenges, we specifically propose a Detection Transformer (DETR) architecture optimized with a Neural Architecture Search (NAS) approach called NAS-DETR for object detection in sonar images. First, an improved Zero-shot Neural Architecture Search (NAS) method based on the maximum entropy principle is proposed to identify a real-time, high-representational-capacity CNN-Transformer backbone for sonar image detection. This method enables the efficient discovery of high-performance network architectures with low computational and time overhead. Subsequently, the backbone is combined with a Feature Pyramid Network (FPN) and a deformable attention-based Transformer decoder to construct a complete network architecture. This architecture integrates various advanced components and training schemes to enhance overall performance. Extensive experiments demonstrate that this architecture achieves state-of-the-art performance on two Representative datasets, while maintaining minimal overhead in real-time efficiency and computational complexity. Furthermore, correlation analysis between the key parameters and differential entropy-based fitness function is performed to enhance the interpretability of the proposed framework. To the best of our knowledge, this is the first work in the field of sonar object detection to integrate the DETR architecture with a NAS search mechanism.

Underwater object detection in sonar imagery with detection transformer and Zero-shot neural architecture search

TL;DR

This work tackles underwater sonar object detection, where low resolution and high noise hinder performance. It introduces NAS-DETR, an entropy-guided zero-shot NAS framework that designs a CNN–Transformer backbone and couples it with a Deformable DETR–style detector, augmented by a content–position decoupled query initialization and a multi-task loss. The approach provides state-of-the-art results on URPC2021 and URPC2022 while preserving real-time speed, and it offers interpretability through Spearman-based analysis linking backbone depth and width to differential entropy. TensorRT quantization further boosts inference speed, making NAS-DETR practical for industrial deployment, with future work focusing on lightweight search, multi-modal fusion, and robust adaptation to dynamic environments.

Abstract

Underwater object detection using sonar imagery has become a critical and rapidly evolving research domain within marine technology. However, sonar images are characterized by lower resolution and sparser features compared to optical images, which seriously degrades the performance of object detection.To address these challenges, we specifically propose a Detection Transformer (DETR) architecture optimized with a Neural Architecture Search (NAS) approach called NAS-DETR for object detection in sonar images. First, an improved Zero-shot Neural Architecture Search (NAS) method based on the maximum entropy principle is proposed to identify a real-time, high-representational-capacity CNN-Transformer backbone for sonar image detection. This method enables the efficient discovery of high-performance network architectures with low computational and time overhead. Subsequently, the backbone is combined with a Feature Pyramid Network (FPN) and a deformable attention-based Transformer decoder to construct a complete network architecture. This architecture integrates various advanced components and training schemes to enhance overall performance. Extensive experiments demonstrate that this architecture achieves state-of-the-art performance on two Representative datasets, while maintaining minimal overhead in real-time efficiency and computational complexity. Furthermore, correlation analysis between the key parameters and differential entropy-based fitness function is performed to enhance the interpretability of the proposed framework. To the best of our knowledge, this is the first work in the field of sonar object detection to integrate the DETR architecture with a NAS search mechanism.
Paper Structure (24 sections, 2 theorems, 39 equations, 9 figures, 12 tables)

This paper contains 24 sections, 2 theorems, 39 equations, 9 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. DETRs
  4. NAS
  5. Porposed NAS-DETR framework
  6. Entropy of CNN-Transformer backbone
  7. NAS Strategy
  8. Decoder for Detectors
  9. Loss Function
  10. Experiments
  11. Dataset and implementation details
  12. Sonar Image Dataset
  13. Experiment settings
  14. NAS
  15. Results on URPC2021 Dataset
  16. ...and 9 more sections

Key Result

LEMMA 1

(Maximum Differential Entropy). Let $x$ be a continuous random variable with probability density function $p(x)$, mean $\mu$, and variance $\sigma^{2}$. The differential entropy of $X$, defined as: satisfies the inequality: with equality if and only if $x$ follows a Gaussian distribution, i.e.,

Figures (9)

  • Figure 1: The proposed NAS-DETR framework for sonar image object detection.
  • Figure 2: Schematic diagram of the two feedforward modes under the Transformer Block and BottleNeck module.
  • Figure 3: Schematic diagram of the proposed architecture search strategy and fitness function calculation.
  • Figure 4: Mutation operation of the CNN-Transformer hybrid backbone network module.
  • Figure 5: Architectural design of the Decoder and Head components.
  • ...and 4 more figures

Theorems & Definitions (2)

  • LEMMA 1
  • LEMMA 2