Single-Stage Underwater Target Detection Based on Feature Anchor Frame Double Optimization Network

Huilin Ge, Yuewei Dai, Zhiyu Zhu, Xu ZangSensors 2022 · OpenReview

TL;DR

This paper introduces UWNet, a novel single-stage underwater object detection network that tackles the challenges of poor visibility and domain shift in underwater environments. By employing a "double optimization" strategy for feature and anchor frames, UWNet significantly improves detection accuracy, especially for small targets, while maintaining efficiency.

What Problem Does This Paper Address?

Underwater environments present unique challenges for computer vision, particularly for object detection. Factors like color shifts, uneven illumination, blurring, and image distortion make it difficult for standard detection algorithms to perform reliably. Furthermore, the "domain offset" between typical training data and underwater imagery reduces the robustness of general-purpose detectors. Existing underwater detection methods often struggle with small targets, require complex two-stage processing, or lack sufficient accuracy. This paper aims to bridge this gap by developing a more accurate and efficient single-stage detector specifically tailored for underwater scenes.

Methods

The core of UWNet lies in its "double optimization" of anchor boxes and features. This is achieved through several key components:

  • Composite Connection Backbone Network: This novel backbone combines multiple existing networks (specifically VGG16 and ResNet50) to leverage their strengths. It fuses features from different levels of abstraction, aiming to enhance feature discrimination and contextual relevance, which is crucial for dealing with the degraded quality of underwater images.
  • ASPP+ Module: An enhanced Atrous Spatial Pyramid Pooling (ASPP) module is integrated. This module uses multiple dilated convolutions with different rates to capture multi-scale contextual information, improving the network's ability to detect objects of varying sizes.
  • Receptive Field Enhancement Module (RFA/RFA+): This module, inspired by ResNet and Inception, uses multi-path parallel branches with different kernel sizes and atrous convolutions to expand the receptive field. This helps in capturing more comprehensive features, especially for small or partially obscured targets.
  • Predictive Optimization Strategies: This is a two-stage refinement process.
  • Preprocessing Stage: Initial classification and regression are performed to filter out obvious background boxes and provide better candidate anchor boxes. An RFA+ module is specifically used here to enhance the detection of small objects.
  • Optimization Stage: This stage further refines the anchor boxes and features through a second round of regression. It addresses feature-anchor misalignment and uses a channel attention mechanism to focus on relevant object features. Unlike previous methods like RefineDet, UWNet uses six prediction layers for refinement, allowing for more detailed adjustments.

Datasets

  • Labeled Fishes in the Wild (LFW): This dataset, provided by the National Oceanic and Atmospheric Administration, was used for training and evaluation. The authors filtered and extended it with annotations for small and ambiguous objects.
  • MS COCO 2015: Used for evaluating UWNet's performance on a more general object detection benchmark.

Key Insights and Ideas

  • The Power of Composite Backbones: Combining existing backbone architectures can unlock greater feature representation capabilities than using a single backbone, especially for challenging domains like underwater imaging.
  • Multi-Scale Context is King: The ASPP+ and RFA modules highlight the importance of capturing features at various scales to effectively detect objects in cluttered and visually degraded underwater scenes.
  • Two-Stage Refinement for Single-Stage Efficiency: UWNet demonstrates that a two-step prediction refinement process, inspired by two-stage detectors, can be integrated into a single-stage framework to significantly boost accuracy without sacrificing too much speed.
  • More Prediction Layers = Better Refinement: Using six prediction layers for refinement, as opposed to the four used in prior work, leads to improved detection performance, particularly on underwater datasets where fine-tuning is critical.

Key Results

  • 80.2 mAP on Labeled Fishes in the Wild: UWNet achieved this impressive result, outperforming most single-stage detectors and even some two-stage methods on this dataset.
  • 2.1 AP Improvement: Compared to a baseline, UWNet achieved a 2.1 AP improvement, attributed to its powerful feature extraction and multi-scale modules.
  • 32.4 AP at 300x300 Input: The model provides a strong accuracy of 32.4 AP even at a moderate input resolution.
  • Improved Small Object Detection: The performance on small objects saw a significant boost, from 10.9 to 15.1 APs.
  • Competitive on MS COCO: UWNet also showed strong performance on the MS COCO dataset, achieving the best results among single-stage detectors in some metrics.

Limitations and Future Work

  • Dataset Specificity: The authors note that the Labeled Fishes in the Wild dataset might not fully represent extreme underwater conditions like severe blur or occlusion, which could limit the observed improvement.
  • Input Size Limitation: The current framework is limited by the input image size, making it difficult to train with very high-resolution images. This restricts its performance potential.
  • Future Directions: The paper suggests exploring 3D surface visualization and potentially using extreme learning techniques for target detection in underwater environments. Further research is needed to overcome the limitations related to input image size and to handle more extreme underwater conditions.

Summary updated

Abstract

Objective: The shallow underwater environment is complex, with problems of color shift, uneven illumination, blurring, and distortion in the imaging process. These scenes are very unfavorable for the reasoning of the detection network. Additionally, typical object identification algorithms struggle to maintain high resilience in underwater environments due to picture domain offset, making underwater object detection problematic. Methods: This paper proposes a single-stage detection method with the double enhancement of anchor boxes and features. The feature context relevance is improved by proposing a composite-connected backbone network. The receptive field enhancement module is introduced to enhance the multi-scale detection capability. Finally, a prediction refinement strategy is proposed, which refines the anchor frame and features through two regressions, solves the problem of feature anchor frame misalignment, and improves the detection performance of the single-stage underwater algorithm. Results: We achieved an effect of 80.2 mAP on the Labeled Fish in the Wild dataset, which saves some computational resources and time while still improving accuracy. On the original basis, UWNet can achieve 2.1 AP accuracy improvement due to the powerful feature extraction function and the critical role of multi-scale functional modules. At an input resolution of 300 × 300, UWNet can provide an accuracy of 32.4 AP. When choosing the number of prediction layers, the accuracy of the four and six prediction layer structures is compared. The experiments show that on the Labeled Fish in the Wild dataset, the six prediction layers are better than the four. Conclusion: The single-stage underwater detection model UWNet proposed in this research has a double anchor frame and feature optimization. By adding three functional modules, the underwater detection of the single-stage detector is enhanced to address the issue that it is simple to miss detection while detecting small underwater targets.
BibTeX
@inproceedings{ge2022single,
  author = {Huilin Ge and Yuewei Dai and Zhiyu Zhu and Xu Zang},
  title = {{Single-Stage Underwater Target Detection Based on Feature Anchor Frame Double Optimization Network}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=3wWMbRFKjx},
}
Figure