AquaFeat+: an Underwater Vision Learning-based Enhancement Method for Object Detection, Classification, and Tracking

Technical Summary: AquaFeat+ for Underwater Vision Tasks

Problem Statement
Underwater robotic applications, such as species identification, debris monitoring, and infrastructure inspection, rely heavily on computer vision systems. However, the underwater environment presents severe challenges, including low lighting, color distortion, light attenuation, and scattering. These factors degrade visual data quality, causing a loss of contrast and sharpness that directly compromises the performance of object detection, classification, and tracking models.

Existing enhancement methods often face two critical limitations:

1. Task Misalignment: Most approaches are optimized for human perceptual quality (aesthetic appeal) rather than the preservation or extraction of features critical for automated analysis. An image that looks visually appealing may still obscure essential patterns, negatively impacting algorithmic performance.
2. Computational Efficiency: Many deep learning-based enhancement models incur high computational demands and long processing times, hindering their applicability in real-time robotic operations.

Methodology: AquaFeat+
The authors propose AquaFeat+, a plug-and-play pipeline designed to enhance features specifically for automated vision tasks rather than human perception. The architecture is trained end-to-end, guided directly by the loss function of the final application (detection, classification, or tracking), ensuring the enhancement is relevant to the downstream task.

The model comprises three primary modules:

1. Color Correction Module: A non-trainable preprocessing step that mitigates chromatic distortions and selective light absorption. It analyzes the average intensities of the Red, Green, and Blue channels and adjusts the color distribution to the median channel, providing a white-balanced, cleaner input for subsequent layers.
2. Feature Enhancement Process: This core process integrates with downstream tasks through three stages:
   • Underwater-Feature Enhancement Network (U-FEN): Processes RGB frames in three parallel streams (original, 1/4, and 1/8 resolution) using shared weights. It replaces the standard first layer with a SpecialConv layer, a trainable, content-aware mechanism that dynamically adjusts contrast based on per-channel statistics (mean and standard deviation). The network utilizes a U-Net architecture with dense skip connections and LeakyReLU activation to prevent vanishing gradients.
   • Global-Scale Attention Module (GSAM): A hybrid architecture designed to refine features via a dual-pathway design. It combines a Global Feature-Aware (GFA) module, which models long-range spatial dependencies, with a Scale-Aware Feature Aggregation (SAFA) mechanism for multi-scale fusion. The outputs are combined with the original input via residual connections to enrich the representation with both global context and multi-scale details.
   • Final Feature Aggregation: The output of the GSAM is resized and concatenated with the lowest-resolution feature stream (1/8 scale). A final 3×3 convolutional layer merges these into a comprehensive representation.
3. Adaptive Residual Output: The final stage generates an enhancement residual map using a terminal SpecialConv layer. This residual is normalized (via tanh) and added to the original input image, preserving structural integrity while applying learned enhancements.

Key Contributions

• Task-Oriented Enhancement: The proposal of AquaFeat+, a module specifically designed to enhance hierarchical features in low-light underwater videos to optimize detection, classification, and tracking performance.
• End-to-End Training: Demonstrating the effectiveness of training the enhancement module directly guided by the final task's loss function, rather than generic image quality metrics.
• Dataset Adaptation: The adaptation of the FishTrack23 dataset with a specific focus on optimizing object detection, classification, and tracking tasks for robotic perception, including data cleaning and class grouping strategies to address imbalance.

Experimental Results
The method was evaluated on the FishTrack23 dataset using YOLOv8m/YOLOv10s for detection/tracking and YOLOv11s-cls for classification. Results were benchmarked against the predecessor (AquaFeat), other enhancement techniques (FeatEnHancer), and baseline models.

• Object Detection: AquaFeat+ (YOLOv8m) achieved the highest F1-Score (0.688), indicating the best balance between precision and recall. While it did not achieve the absolute highest mAP0.5 (0.556 vs. 0.557 for AquaFeat YOLOv10s), it maintained extremely competitive performance while significantly improving recall over the baseline YOLOv8m (0.624 vs. 0.582).
• Classification: AquaFeat+ outperformed all other approaches in precision (0.816), recall (0.791), and F1-Score (0.791). It demonstrated a consistent ability to identify and balance positive predictions, surpassing existing methods like FeatEnHancer and the baseline YOLOv11s-cls.
• Tracking: In multi-object tracking, AquaFeat+ (YOLOv10s) achieved the highest HOTA score (55.206) and the top AssA score (60.194), indicating a superior balance between detection and association accuracy. It also secured the second-best IDF1 score (68.094), demonstrating effective long-term identity consistency.

Qualitative analysis confirmed that AquaFeat+ was often the only method capable of correctly detecting partially occluded objects at image borders or identifying fish in low-visibility scenarios where baseline models and other enhancement techniques failed.

Significance and Claims
The paper claims that AquaFeat+ validates its ability to significantly elevate perception capabilities in complex underwater scenes. By integrating feature enhancement directly into the vision task pipeline and optimizing for downstream performance rather than visual appeal, the method addresses the gap between image enhancement and automated analysis. The results suggest that AquaFeat+ is a flexible, efficient solution that can be seamlessly integrated into diverse robotic platforms to improve the reliability of scene analysis for decision-making. The authors note that future work will involve extending the application to tasks such as depth estimation and semantic segmentation, as well as evaluating the model on new real-world datasets.