Scholarlink

<p>Deep learning methods have significantly advanced the semantic segmentation of high-resolution remote sensing images. However, challenges such as suboptimal accuracy, slow convergence, and blurred object boundaries persist, particularly in complex urban environments. To address these issues, we propose an enhanced architecture called Inception Attention Residual U-Net (IARU-Net), which integrates inception modules, attention mechanisms, and residual connections into the U-Net backbone. This design improves multi-scale feature extraction, enhances boundary delineation, and preserves fine spatial details. Additionally, we introduce a novel hybrid hyperparameter optimization algorithm, PSOLFGA, which combines Particle Swarm Optimization (PSO), Levy Flight (LF), and Genetic Algorithm (GA). PSOLFGA accelerates convergence, improves search diversity, and avoids local optima. Experimental results demonstrate that the proposed IARU-Net, when optimized using PSOLFGA, consistently outperforms baseline models and individual optimization variants across multiple evaluation metrics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An optimized multi-objective-derived U-net for semantic segmentation in high-resolution remote sensing images

  • 期刊论文

  • 发表日期：

    2025年6月11日

查看全文
  • Nagamani Gonthina,
  • L V Narasimha Prasad

摘要

Deep learning methods have significantly advanced the semantic segmentation of high-resolution remote sensing images. However, challenges such as suboptimal accuracy, slow convergence, and blurred object boundaries persist, particularly in complex urban environments. To address these issues, we propose an enhanced architecture called Inception Attention Residual U-Net (IARU-Net), which integrates inception modules, attention mechanisms, and residual connections into the U-Net backbone. This design improves multi-scale feature extraction, enhances boundary delineation, and preserves fine spatial details. Additionally, we introduce a novel hybrid hyperparameter optimization algorithm, PSOLFGA, which combines Particle Swarm Optimization (PSO), Levy Flight (LF), and Genetic Algorithm (GA). PSOLFGA accelerates convergence, improves search diversity, and avoids local optima. Experimental results demonstrate that the proposed IARU-Net, when optimized using PSOLFGA, consistently outperforms baseline models and individual optimization variants across multiple evaluation metrics.