光学遥感图像显著目标检测的边缘增强轻量型模型探究

Exploration of a Lightweight Model for Edge Enhancement for Salient Object Detection in Optical Remote Sensing Images

摘要: 光学遥感图像(ORSI)场景复杂且目标尺度多变，现有显著目标检测(SOD)模型存在参数量大、计算成本高或边缘信息利用不足导致目标边缘模糊等问题。针对上述缺陷，提出一种边缘引导的轻量型注意力网络模型(EGLANet)，实现光学遥感图像显著目标的高精度、轻量化检测。该模型以MobileNet V3为骨干网络生成五级初级特征图，先通过由跨层相关性模块(CLC)和多尺度并行卷积构成的边缘多尺度注意力模块(EMSAM)提取四级边缘特征图与精细边缘图，再利用CLC、融合1 × k/k × 1分组卷积的多尺度注意力模块(MSAM)结合边缘引导模块(EGM)，将边缘信息融入目标检测过程以精准定位显著目标；损失函数采用二值交叉熵(BCE)与交并比(IoU)损失的组合形式，同时约束边缘提取与目标检测任务。在公开光学遥感图像显著目标检测数据集EORSSD上开展对比实验，选取10种先进模型作为对比对象，从定性与定量角度验证模型性能。实验结果表明，EGLANet实现了参数量仅为1.75 M，显著目标检测效果却优于其它模型，在保证检测精度的同时提升计算效率，为光学遥感图像显著目标检测提供了一种高效的轻量化解决方案。

Abstract: Optical Remote Sensing Images (ORSI) feature complex scenes and variable target scales. Existing Salient Object Detection (SOD) models suffer from issues such as large parameter count, high computational cost, or insufficient utilization of edge information, leading to blurred target edges. To address these shortcomings, we propose an Edge-Guided Lightweight Attention Network (EGLANet) to achieve high-precision and lightweight detection of salient objects in optical remote sensing images. This model utilizes MobileNet V3 as the backbone network to generate five-level primary feature maps. It first extracts four-level edge feature maps and fine edge maps through an Edge Multi-Scale Attention Module (EMSAM) composed of a Cross-Layer Correlation (CLC) module and multi-scale parallel convolution. Then, it integrates edge information into the target detection process using a combination of CLC, a multi-scale attention module (MSAM) that fuses 1 × k/k × 1 grouped convolution, and an Edge Guidance Module (EGM) to accurately locate salient objects. The loss function combines Binary Cross Entropy (BCE) and Intersection over Union (IoU) losses, simultaneously constraining edge extraction and target detection tasks. Comparative experiments were conducted on the public optical remote sensing image salient object detection dataset EORSSD, selecting 10 advanced models as benchmarks to verify the model performance from both qualitative and quantitative perspectives. The experimental results show that EGLANet achieves superior salient object detection performance with only 1.75 M parameters, improving computational efficiency while maintaining detection accuracy, providing an efficient and lightweight solution for salient object detection in optical remote sensing images.