Gamma-Asymmetric Enhancement Module

• Gamma-Asymmetric Enhancement Module is a deep learning feature processing unit that integrates asymmetric and dilated convolutions with learnable gamma scaling to recover fine-grained semantic details.
• It employs a four-branch architecture that progressively fuses multi-scale information with up-sampling and channel attention to enhance object localization in challenging underwater environments.
• Empirical studies on underwater camouflaged object detection reveal that GAE improves the weighted F-measure by 2.6% and reduces MAE significantly, showcasing its practical impact.

Gamma-Asymmetric Enhancement (GAE) Module is a feature processing unit developed to strengthen multi-scale representations, recover fine-grained semantic detail, and inject adaptive contextual weighting within deep convolutional architectures. It was introduced as a central mechanism in the Semantic Localization and Enhancement Network (SLENet) for Underwater Camouflaged Object Detection (UCOD), a domain noted for severe optical distortions, blurred boundaries, and low-contrast textures. GAE specializes in integrating asymmetric and dilated convolutional operations with end-to-end learnable gamma-style scaling, producing refined features for challenging visual scenarios (Wang et al., 4 Sep 2025).

1. Architectural Formulation

In SLENet, GAE operates on the four hierarchical features $\{X_i, i \in \{1,2,3,4\}\}$ extracted by a SAM2 encoder with lightweight adapters. Each feature $X_i$ is processed by a dedicated GAE block comprising four sequential branches $r=1,2,3,4$. The design within each branch is characterized by:

• Initial 1×1 convolution for channel compression,
• A stack of asymmetric convolution and max-pooling layers for directional selectivity and parameter efficiency,
• Dilated convolution ($\text{dilation}=2$) for expanded receptive field.

Branches dependencies are encoded such that, starting with $r=2$, each branch fuses the up-sampled output of its predecessor to incorporate higher-resolution cues. After all branches complete, channel attention and learnable scaling $\gamma$ are applied to the deepest branch’s output, yielding an enhanced feature $F_i$ with the same spatial and channel dimensions as the input.

2. Mathematical Specification

The flow of computation within each GAE block is rigorously formalized:

• For branch $r=1$:

$$D_i^1 = C_{\mathrm{dil}}\bigl(\mathrm{AMP}_{\times3}(X_i^1)\bigr)$$

• For branches $r=2,3,4$:

$$C_i^r = C_{\mathrm{asy}}\bigl(\text{Cat}\bigl(X_i^r,\;\text{Up}(D_i^{r-1}, X_i^r)\bigr)\bigr)$$

$$D_i^r = C_{\mathrm{dil}}\bigl(\mathrm{AMP}_{\times(4-r)}(C_i^r)\bigr)$$

• Final aggregation, channel attention, and scaling:

$F_i = \gamma\,\bigl(D_i^4 \otimes CA(D_i^4)\bigr)$

where $C_{\text{asy}}$ indicates asymmetric convolution, $C_{\text{dil}}$ is dilated convolution, $\mathrm{AMP}_{\times n}$ is a stack of $n$ asymmetric-max-pool pairs, $Cat$ is channel concatenation, $Up$ is spatial up-sampling, $CA$ is channel attention, $\otimes$ is element-wise multiplication, and $\gamma$ is a learnable scalar.

3. Design Motivations

Three main design rationales drive GAE’s architecture:

• Asymmetric Convolutions: By decomposing a $k\times k$ kernel into $k\times 1$ and $1\times k$ filters, the module sharply reduces parameter count—essential when deploying on computationally demanding backbones. Moreover, this structure directly enhances sensitivity to anisotropic textures, especially horizontal or vertical alignments prevalent in camouflaged marine contours.
• Dilated Convolutions: Employing dilation rate $2$ allows GAE to expand receptive field coverage without increasing stride or reducing spatial resolution, facilitating foreground-background separation in ambiguous, low-contrast underwater images.
• Progressive Multi-Branch Fusion: The four-branch topology ensures the fusion of both local and global cues. Early branches capture fine textures; succeeding branches integrate context with increased resolution through up-sampled cross-branch fusion.
• Learnable Gamma Scaling: The scalar $\gamma$ dynamically calibrates feature intensity, functioning analogously to gamma correction. This allows channel-wise adaptation and contrast normalization as learned during training—a property valuable for underwater data exhibiting highly variable contrast and brightness.

4. Functional Role in SLENet Pipeline

The GAE module is integrated into two principal sub-networks of SLENet—Localization Guidance Branch (LGB) and Multi-Scale Supervised Decoder (MSSD):

• Localization Guidance Branch: For each fused feature at level $i$, LGB applies GAE post-fusion:

$F_2^l = GAE(Down(X_1^l) \oplus X_2^l)$

$$F_i^l = GAE(Down(F_{i-1}^l) \oplus X_i^l), \; i=3,4$$

where $X_i^l$ is a 1×1 compressed backbone feature, $Down(\cdot)$ is spatial down-sampling, and $\oplus$ is element-wise addition. The terminal output $F_4^l$ is processed by convolution to yield a coarse localization map $M$.

• Multi-Scale Supervised Decoder: GAE-refined features $F_i$ are fused with up-sampled outputs from deeper decoder layers. This composite is passed through spatial attention, residual connections, and final 1×1 convolution to produce segmentation logits $P_i$.

5. Training Protocols and Optimization

GAE’s parameters, alongside those of LGB, MSSD, and adapters, are learned with all SAM2 backbone weights frozen. Optimization utilizes AdamW (initial learning rate $5\times10^{-4}$, cosine decay), with inputs of size $352 \times 352$, batch size $16$, and $100$ epochs. Supervisory signals comprise weighted binary cross-entropy (BCE) and weighted Intersection-over-Union (IoU) losses targeting each segmentation logit $P_i$, and separate BCE loss on the localization map $M$ weighted by a factor $\omega_m$:

$$\omega_m = \max\left(\mu\left(1-\frac{epoch}{epochs}\right),\,0.1\right)$$

where $\mu=0.6$ and the weight linearly decays to $0.1$. The gamma parameter $\gamma$ for each GAE block is optimized end-to-end, with no manual tuning required.

6. Empirical Performance and Qualitative Impact

Empirical results from ablation studies on the DeepCamo dataset demonstrate that GAE provides substantial gains:

Configuration	Weighted-F ($F^w_\beta$)	MAE
Baseline (no GAE)	0.764	0.026
GAE only	0.784 (+2.6%)	0.023
GAE + LGB + MSSD	0.800	0.022

Insertion of GAE alone yields a +2.6% absolute increase in weighted F-measure and an ≈11.5% relative drop in MAE. When fully integrated with SLENet's other modules, it sets state-of-the-art performance benchmarks. Qualitative analyses illustrate the module’s strength in recovering thin anatomical structures and preserving precise boundaries amidst blur and variable contrast, attributes intrinsic to natural underwater camouflage (Wang et al., 4 Sep 2025).

7. Contextual Significance in Underwater Camouflaged Object Detection

GAE’s architectural choices and adaptive mechanisms are particularly attuned to the demands of underwater camouflaged object detection—a domain where the objects of interest are often indistinguishable from background, dominated by low SNR textures, blended outlines, and anisotropic patterns. The asymmetric and dilated filtering, multi-branch aggregation, and dynamic gamma scaling collectively address the need for robust, multi-scale, context-aware feature enhancement. A plausible implication is that such module designs could generalize to other visual recognition tasks exhibiting similar challenges of contour ambiguity, multi-resolution context, and photometric instability.