AquaFeat+ Pipeline for Underwater Vision

• AquaFeat+ is a plug-and-play, learning-based framework that enhances underwater visual features for detection, classification, and tracking.
• It integrates a deterministic color correction module, multi-scale feature enhancement via U-FEN with attention fusion, and an adaptive residual output.
• End-to-end training under downstream task loss yields robust performance improvements on the FishTrack23 dataset under challenging underwater conditions.

AquaFeat+ is a learning-based, plug-and-play enhancement pipeline designed for underwater visual perception tasks. By targeting the improvement of feature representations directly relevant to object detection, classification, and tracking, AquaFeat+ addresses the substantial challenge posed by underwater image degradation arising from low lighting, color distortion, and turbidity. Distinct from generic perceptually-motivated image enhancement, the pipeline is trained end-to-end strictly under the loss of the downstream task, thus optimizing its transformation to serve automated computer vision objectives, as substantiated in large-scale experiments on the FishTrack23 dataset (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

1. Pipeline Architecture and Data Flow

AquaFeat+ comprises three sequential, modular stages:

1. Color Correction Module (CCM): A non-trainable, histogram-based white-balance transform that rapidly normalizes per-channel intensities.
2. Hierarchical Feature Enhancement: Multi-scale enhancement via the Underwater-Feature Enhancement Network (U-FEN) with shared weights, augmented by attention-based fusion mechanisms.
3. Adaptive Residual Output (ARO): Application of a learned, channel-wise residual—normalized by tanh—mapped back to the color-corrected image.

The operational sequence is as follows:

• Input image $I \in \mathbb{R}^{H \times W \times 3}$
• CCM: $I \to I_{cc}$
• Multi-scale branch: $I_{cc}$ downsampled to $1\times$, $1/4\times$, $1/8\times$; each passes through U-FEN (shared weights), producing $[f_1, f_2, f_3]$
• Feature fusion (SAFA and/or GSAM depending on sub-version): Deep attention and aggregation yielding fused feature $F$
• ARO: SpecialConv + tanh applied, producing enhancement $R$, and output $I_{enh} = I_{cc} + R$ (or $I \to I_{cc}$0 in some YOLO integrations)
• Enhanced image $I \to I_{cc}$1 directly feeds the detection/classification/tracking backbone (e.g., YOLOv8m)

Illustratively, AquaFeat+ functions seamlessly as a pre-task enhancer across object detection (YOLO), classification (YOLOv11s-cls), and tracking (ByteTrack with YOLO boxes) (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

2. Core Module Details and Mathematical Formulation

2.1 Color Correction Module (CCM)

This deterministic preprocessing rescale brings $I \to I_{cc}$2 to comparable dynamic range. Explicitly:

Given channel means

$I \to I_{cc}$3

for $I \to I_{cc}$4, set $I \to I_{cc}$5. For each channel:

$I \to I_{cc}$6

No gradient flows through CCM; it is excluded from the trainable pipeline (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

2.2 Underwater-Feature Enhancement Network (U-FEN)

For scales $I \to I_{cc}$7, $I \to I_{cc}$8, $I \to I_{cc}$9:

• Input $I_{cc}$0 in $I_{cc}$1
• SpecialConv first layer: computes per-channel mean and std, projects via a small MLP to generate modifiers $I_{cc}$2 for dynamic convolution
• Encoder stack: six $I_{cc}$3 convolutions with dense skip connections; all activations LeakyReLU ($I_{cc}$4); channels fixed at $I_{cc}$5

Formally, the first U-FEN operation:

$I_{cc}$6

where $I_{cc}$7 denotes convolution and $I_{cc}$8 depends on statistics of $I_{cc}$9.

2.3 Fusion and Attention: SAFA and GSAM

• Scale-Aware Feature Aggregation (SAFA):
  • Features from two scales are mapped into a common key/query/value space via $1\times$0 convolutions. Multi-head dot-product attention is computed and the result upsampled and fused, followed by $1\times$1 convolution.

Example (at downsampled spatial resolution):

$1\times$2

Multi-head:

$1\times$3

Concatenate results and upsample.

• Global-Scale Attention Module (GSAM):
  • Incorporates both global (self-attention on lowest-resolution features) and scale-aware fusion (from multiple resolutions). Final aggregation involves summing raw low-res features, global attended features, and inter-scale fusion, then upsampling and channel-concatenation.

2.4 Adaptive Residual Output (ARO)

The final output is a residual map $1\times$4 generated by a SpecialConv followed by a tanh normalization:

$1\times$5

Then the enhanced image is computed as $1\times$6 or $1\times$7 (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

3. End-to-End Training Paradigm

AquaFeat+ is trained strictly under the downstream task’s loss—no perceptual or low-level image losses are introduced.

• For detection/tracking (YOLOv8m backbone), loss is:

$1\times$8

with $1\times$9: cross-entropy; $1/4\times$0: (C)IoU-based regression; $1/4\times$1: objectness BCE.

• For classification:

$1/4\times$2

Gradients flow through all trainable enhancement modules, adapting enhancement strategy explicitly for object features most relevant to the detection/classification head.

A plausible implication is that because the only supervision comes from the vision task’s detection/classification loss, the enhancement prioritizes semantically-relevant features over generic photometric criteria (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

4. Implementation, Hyper-parameters, and Resource Profiling

Key implementation configurations:

Parameter	Value / Setting
Feature channels	64 (all U-FEN and attention modules)
Multi-scale	1×, 1/4×, 1/8×
Attention heads	8 (SAFA/GSAM)
Activations	LeakyReLU($1/4\times$3); tanh (final)
Optimizer	AdamW (lr = $1/4\times$4 or $1/4\times$5)
Batch size	6 (detection/tracking), 16–32 (classification)
Epochs	50 (detection), 30 (classification)
Hardware	RTX 4070 Ti, Ryzen 7 8700G, 32 GB RAM
Data augmentation	Horizontal flip, scale jitter, color jitter

The additional parameter count for AquaFeat+ is ≈1.7M, total $1/4\times$6M with YOLOv8m backbone. Throughput is 46.5 FPS on RTX 4070 Ti for detection (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

5. Empirical Performance: Detection, Classification, Tracking

Aquafeat+ has been benchmarked extensively on the FishTrack23 dataset, which consists of annotated underwater videos with object (fish) bounding boxes and taxonomy, spanning >14,000 test frames.

Detection:

Method	Prec.	Rec.	F1	[email protected]	[email protected]:0.95
YOLOv8m	0.792	0.582	0.677	0.528	0.319
FeatEnHancer + YOLOv8m	0.753	0.582	0.657	0.515	0.293
AquaFeat + YOLOv8m	0.767	0.624	0.688	0.556	0.332
AquaFeat+ (YOLOv8m)	N/R	N/R	N/R	N/R	N/R

Additional specialized experiments (Silva et al., 17 Aug 2025) (using a different split) report:

Method	[email protected]	[email protected]:0.95	Precision	Recall	FPS
YOLOv8m (baseline)	0.647	0.387	0.847	0.584	321.5
AquaFeat+ + YOLOv8m	0.677	0.421	0.877	0.624	46.5

Classification (YOLOv11s-cls backbone):

Method	Prec.	Rec.	Acc.	F1
YOLOv11s-cls	0.723	0.764	0.764	0.737
FeatEnHancer+	0.746	0.779	0.779	0.752
AquaFeat+	0.816	0.791	0.791	0.791

Tracking (ByteTrack + YOLO):

Method	HOTA	MOTA	DetA	AssA	IDF1
YOLOv8m	52.75	53.78	51.42	54.41	65.10
FeatEnHancer+	47.48	37.23	41.15	54.97	59.42
AquaFeat+ (YOLOv10s)	55.21	55.01	50.90	60.19	68.09

AquaFeat+ achieves highest or near-highest scores across nearly all metrics, particularly excelling in F1 and HOTA (Silva et al., 14 Jan 2026). Impactful qualitative differences include more robust detection of occluded or low-contrast targets, fewer species misclassifications, and greater ID consistency across challenging tracking scenes.

6. Comparative and Ablation Analysis

Comparative evaluation demonstrates clear incremental improvements through the AquaFeat, AquaFeat+, and baseline progression. The introduction of attention-based fusion (GSAM/SAFA) and adaptive residual output contributes the largest accuracy jumps, especially in F1 (detection/classification) and HOTA (tracking).

Ablation studies within controlled internal experiments indicate that disabling attention fusion (SAFA) results in a typical $1/4\times$7–$1/4\times$8 point [email protected] drop, and removal of the adaptive residual output reduces mAP by approximately $1/4\times$9 points (Silva et al., 17 Aug 2025). This suggests critical importance of both multi-scale feature integration and explicit residual correction.

7. Application Scope and Deployment Feasibility

AquaFeat+ is hardware efficient—runtime of 46.5 FPS allows real-time operation on modern GPUs. It introduces a modest parameter overhead (~1.7M) relative to typical modern detection/classification backbones, and supports further acceleration through mixed-precision and potential pruning.

The design is agnostic to target backbone, enabling integration with various detectors (YOLOv8m, YOLOv10s), classifiers (YOLOv11s-cls), and trackers (ByteTrack); this plug-and-play nature makes it suitable for deployment in marine monitoring, infrastructure inspection, and any robotic setting requiring underwater visual perception (Silva et al., 17 Aug 2025, Silva et al., 14 Jan 2026).

A plausible implication is that the task-driven, end-to-end learning regimen could be extended or adapted to other vision domains where signal degradation impedes feature robustness.