西安石油大学学报（自然科学版）2025，Vol.40Issue(5)：133-142,10.DOI:10.3969/j.issn.1673-064X.2025.05.016
基于集成学习的双分支混合注意力块去雾算法
Double-Branch Hybrid Attention Block Image Dehazing Algorithm Based on Ensemble Learning
摘要
Abstract
A dual-branch hybrid attention block dehazing network based on ensemble learning is proposed to address the issues of color distortion and incomplete dehazing caused by existing deep learning methods when processing real foggy images.The network is com-posed of a global feature extraction branch and a high-frequency texture feature extraction branch,and a tail integrated learning fusion is applied.The global feature extraction subnet uses Res2Net as the pre-training model to initialize weights and improve the robustness of global feature extraction.The high-frequency texture feature extraction subnet uses U-Net as the main network,combining Mixed Atten-tion blocks,SKFusion Layer,and Soft Recon,to make this branch more able to focus on high-frequency texture features in the current dataset.Finally,two branches are fused and mapped into a clear image.Compared with state-of-the-art methods,the proposed algorithm achieves good visual effects on the RESIDE synthetic dataset,and the objective metrics on the NTIRE2020 and NTIRE2021 datasets show an improvement of 12.09 dB in PSNR and 40.7%in SSIMover the latest DehazeFormer defogging model.关键词
图像去雾/注意力机制/集成学习/编码器-解码器Key words
image dehazing/attention mechanism/integrated learning/encoder-decoder分类
信息技术与安全科学引用本文复制引用
张仙伟,陈泽锐,任帅..基于集成学习的双分支混合注意力块去雾算法[J].西安石油大学学报（自然科学版）,2025,40(5):133-142,10.基金项目
陕西省自然科学基金(2020JM-543)(非线性自适应图正则的子空间聚类算法研究) （2020JM-543）
陕西省重点研发计划2020GY-038(基于信息物理融合的智能油田监控系统关键技术研究),2021GY-083(基于深度学习双目智能感知与定位技术研究) （基于信息物理融合的智能油田监控系统关键技术研究）
