UKANFormer Decoder: Robust Coral Segmentation

• UKANFormer Decoder is a specialized component for semantic segmentation that combines convolutional and transformer-based techniques to handle noisy supervision effectively.
• It employs the GL-Trans block to fuse local detail recovery with global semantic coherence, ensuring precise boundary delineation and structural continuity.
• Empirical results demonstrate significant gains over baseline models, with improvements of over 2 percentage points in coral-class IoU and pixel accuracy despite coarse label inputs.

The UKANFormer decoder is a specialized architectural component designed for robust semantic segmentation under noisy supervision, with primary application in high-precision coral reef mapping. It is distinguished by the integration of the Global-Local Transformer (GL-Trans) block, which fuses convolutional and self-attention mechanisms to address both global semantic coherence and fine-grained local detail recovery. This design enables UKANFormer to outperform traditional models in scenarios characterized by coarse or unreliable labels, producing high-fidelity segmentation maps that surpass even the quality of their training data.

1. Design Objectives and Context

The UKANFormer decoder addresses two principal challenges endemic to coral reef segmentation with noisy supervision: (1) rectification of global semantic discontinuities and fragmented connectivity caused by imprecise labels, and (2) enhancement of local boundary accuracy and fine detail, particularly where label boundaries are ambiguous. The decoder operates downstream of a Kolmogorov-Arnold Network-based encoder, leveraging nonlinear feature extraction for input representations, which the decoder must further process to yield semantically and spatially accurate segmentation outputs. These design considerations are motivated by the limitations observed in global mapping products such as the Allen Coral Atlas, especially regarding the need for fine boundary delineation and structural continuity.

2. Architecture of the GL-Trans Block

At the core of the decoder is the Global-Local Transformer (GL-Trans) block, a dual-path module comprising:

• A local branch utilizing convolutional operations for enhancement of spatial detail and boundary precision.
• A global branch implementing Transformer-style self-attention for the modeling of long-range dependencies and semantic structure.

2.1 Local Branch

The local branch operates on a feature tensor $X \in \mathbb{R}^{C \times H \times W}$ and involves successive convolution and normalization operations:

• $1\times1$ convolution for semantic integration:

$$F_{1\times1}(X) = \mathrm{BatchNorm}(W_{1\times1} * X)$$

• $3\times3$ convolution for capturing textural and boundary information:

$$F_{3\times3}(X) = \mathrm{BatchNorm}(W_{3\times3} * X)$$

• The local feature representation is obtained by summation:

$$F_{\text{local}} = F_{1\times1}(X) + F_{3\times3}(X)$$

2.2 Global Branch

The global path projects spatial features for attention-based modeling:

• Flatten $X$ to $X_{\text{seq}} \in \mathbb{R}^{L \times D}$, where $L = H \times W$, $D$ as specified by downstream projection.
• Compute query, key, and value matrices:

$$Q = W_Q \cdot X_{\text{seq}}, \quad K = W_K \cdot X_{\text{seq}}, \quad V = W_V \cdot X_{\text{seq}}$$

• Apply scaled self-attention (with learnable bias $\theta \in \mathbb{R}^{L \times L}$):

$$A = \mathrm{Softmax}\left(\frac{QK^T}{\sqrt{D_k}} + \theta\right)$$

• Aggregate global features and restore spatial shape:

$F_{\text{global}} = \mathrm{Reshape}(AV)$

2.3 Fusion and Refinement

The outputs of the two branches are fused and refined:

• Add local and global features, then apply a depthwise separable convolution:

$$F_{\text{DW}} = \mathrm{DWConv}(F_{\text{local}} + F_{\text{global}})$$

• Perform final $1\times1$ convolution with batch normalization to yield decoder output:

$$F_{\text{out}} = \mathrm{BatchNorm}(W_{1\times1} * F_{\text{DW}})$$

This process combines local detail retention with global structural enforcement in a computationally efficient manner due to the depthwise operation.

3. Integration and Role in Semantic Segmentation

The fusion of global and local information in the GL-Trans block explicitly targets the correction of two error sources associated with noisy label supervision: (1) global region connectivity and semantic coherence, and (2) local boundary sharpness and morphological accuracy. The global self-attention path supports contextual filling and correction of fragmented or inconsistent areas, while the convolutional local path ensures edge precision and recovery of high-frequency detail. This combined approach enables the decoder to reconstruct both fine and large-scale structures, even when trained with ambiguous ground truth.

4. Empirical Performance and Validation

Experiments with UKANFormer under both noisy (Allen Coral Atlas) and expert annotation scenarios demonstrate significant performance gains:

Model	Coral-class IoU	Pixel Accuracy
UKANFormer	67.00%	83.98%
UKAN (abl.)	<64.58%	<81.88%

Note: The “<” symbol indicates the baseline's lower performance, by over 2 percentage points in both IoU and pixel accuracy (see Table 2 in the cited work).

Visualizations show that UKANFormer generates more continuous, structurally accurate, and less fragmented segmentations, with enhanced boundary fidelity relative to both noisy labels and model baselines. These results empirically demonstrate that UKANFormer’s predictions can exceed the spatial and semantic quality of the labels used in supervision.

5. Impact of Decoder Design and Ablation Insights

Ablation studies confirm that the incorporation of the GL-Trans block is not only impactful but essential for the observed improvements in robustness and accuracy: “the addition [of GL-Trans] improves the coral-class IoU by 2.42 percentage points and the pixel accuracy by 2.10 percentage points under expert-labeled evaluation ... These improvements confirm that the GL-Trans block ... enhances long-range structure perception and mitigates boundary fragmentation and connectivity breaks...” The dual-path approach, and especially the explicit global-local feature fusion and depthwise refinement, is fundamental to the ability of the decoder to learn discriminative mapping under imperfect supervision and within practical computational constraints.

6. Theoretical and Practical Implications

The UKANFormer decoder challenges the assumption that model performance is rigidly limited by label quality. By leveraging architectural innovation—notably the dual-path GL-Trans block—UKANFormer demonstrates that careful design can mitigate detrimental effects of label noise and provides scalable segmentation where annotated data is scarce or unreliable. This capability is of direct practical relevance to ecological monitoring and other domains where large-scale, high-precision mapping is required but exhaustive expert annotation is infeasible.

In summary, the UKANFormer decoder exemplifies an advanced design for noise-robust semantic segmentation, efficiently synthesizing convolutional and Transformer-based operations to transcend the constraints of label quality and deliver state-of-the-art performance in real-world, imperfect supervision regimes.