<p>Accurate segmentation of small and sparse objects in images remains a fundamental challenge in computer vision and pattern recognition, characterized by three persistent difficulties: severe class imbalance with foreground regions occupying less than 0.5% of the image area, ambiguous object boundaries exhibiting low contrast against surrounding background, and the simultaneous requirement for fine-grained local texture and broad global context modeling across wide scale variations. Existing deep learning segmentation methods address these challenges only partially and in isolation, lacking a unified framework that systematically resolves all three difficulties in concert. To address these issues, we propose EPLGNet (Edge Prior and Local-Global Decoupled Network), an end-to-end deep learning framework for precise segmentation of small and sparse objects. EPLGNet incorporates an Edge Prior Module (EPM) at the encoder-decoder bottleneck, which employs learnable directional Sobel convolutional kernels to explicitly extract boundary gradient responses and embeds precise spatial boundary constraints into the training process via an edge auxiliary supervision signal. In parallel, a Local-Global feature decoupling dual-branch is designed: the local branch captures fine-grained multi-scale textures through parallel dilated convolutions with dilation rates <InlineEquation ID="IEq1"><EquationSource Format="TEX">$\{1,2,4,8\}$</EquationSource><EquationSource Format="MATHML"><math><mrow><mo stretchy="false">{</mo><mn>1</mn><mo>,</mo><mn>2</mn><mo>,</mo><mn>4</mn><mo>,</mo><mn>8</mn><mo stretchy="false">}</mo></mrow></math></EquationSource></InlineEquation>, while the global branch models global contextual dependencies via a multi-resolution pooling pyramid; both branches are adaptively fused through channel attention to achieve explicit decoupling and dynamic integration. At each decoder level, attention gates (AG) and CBAM dual-attention modules are deployed for progressive feature refinement, complemented by a composite loss function combining BCE, Dice, Focal, and boundary-aware losses together with multi-level deep supervision, systematically addressing the extreme class-imbalance problem. Experiments on the publicly available Pancreas CT Segmentation dataset comprising 43,480 two-dimensional CT slices, which instantiates all three challenges in their most demanding form, demonstrate that EPLGNet outperforms six mainstream comparison methods across all six evaluation metrics. Compared with the strongest baseline UCTransNet, EPLGNet achieves a DSC improvement of 3.09 percentage points (0.6821 vs. 0.6512), an IoU improvement of 3.40 percentage points (0.5193 vs. 0.4853), and reduces HD95 from 7.12 mm to 5.34 mm. Ablation studies quantitatively validate the independent contribution of each component, and size-stratified analysis further confirms the particular value of EPM for the micro-scale object subgroup (<5 mm). These results demonstrate that the boundary-aware, locally fine, globally semantic collaborative design paradigm proposed by EPLGNet provides an effective solution for precise automatic segmentation of small and sparse objects on benchmarks featuring extreme foreground sparsity.</p>

EPLGNet: An edge prior and local-global decoupled network for precise small object segmentation under extreme class imbalance

期刊论文
出版模式：

开放获取
发表日期：

2026年6月19日

查看全文

Yuxin Liang,
Haixin Huang,
Yuhao Su,
Deyuan Zhong,
Hongtao Yan,
Yahui Chen,
Shuoshuo Ma,
Qinyan Yang,
Xiaolun Huang

摘要

Accurate segmentation of small and sparse objects in images remains a fundamental challenge in computer vision and pattern recognition, characterized by three persistent difficulties: severe class imbalance with foreground regions occupying less than 0.5% of the image area, ambiguous object boundaries exhibiting low contrast against surrounding background, and the simultaneous requirement for fine-grained local texture and broad global context modeling across wide scale variations. Existing deep learning segmentation methods address these challenges only partially and in isolation, lacking a unified framework that systematically resolves all three difficulties in concert. To address these issues, we propose EPLGNet (Edge Prior and Local-Global Decoupled Network), an end-to-end deep learning framework for precise segmentation of small and sparse objects. EPLGNet incorporates an Edge Prior Module (EPM) at the encoder-decoder bottleneck, which employs learnable directional Sobel convolutional kernels to explicitly extract boundary gradient responses and embeds precise spatial boundary constraints into the training process via an edge auxiliary supervision signal. In parallel, a Local-Global feature decoupling dual-branch is designed: the local branch captures fine-grained multi-scale textures through parallel dilated convolutions with dilation rates $\{1,2,4,8\}$ ${1, 2, 4, 8}$ , while the global branch models global contextual dependencies via a multi-resolution pooling pyramid; both branches are adaptively fused through channel attention to achieve explicit decoupling and dynamic integration. At each decoder level, attention gates (AG) and CBAM dual-attention modules are deployed for progressive feature refinement, complemented by a composite loss function combining BCE, Dice, Focal, and boundary-aware losses together with multi-level deep supervision, systematically addressing the extreme class-imbalance problem. Experiments on the publicly available Pancreas CT Segmentation dataset comprising 43,480 two-dimensional CT slices, which instantiates all three challenges in their most demanding form, demonstrate that EPLGNet outperforms six mainstream comparison methods across all six evaluation metrics. Compared with the strongest baseline UCTransNet, EPLGNet achieves a DSC improvement of 3.09 percentage points (0.6821 vs. 0.6512), an IoU improvement of 3.40 percentage points (0.5193 vs. 0.4853), and reduces HD95 from 7.12 mm to 5.34 mm. Ablation studies quantitatively validate the independent contribution of each component, and size-stratified analysis further confirms the particular value of EPM for the micro-scale object subgroup (<5 mm). These results demonstrate that the boundary-aware, locally fine, globally semantic collaborative design paradigm proposed by EPLGNet provides an effective solution for precise automatic segmentation of small and sparse objects on benchmarks featuring extreme foreground sparsity.