多层级自适应融合的建筑图像检索网络研究OA
Multi-level Adaptive Fusion Network for Building Image Retrieval
针对建筑图像检索中特征提取困难、细节信息丢失、特征融合不充分等问题,提出了一种多层级自适应融合的建筑图像检索网络.首先,在 ResNet50 骨干网络中引入非对称卷积块来提高对目标建筑区域的关注,设计全局信息感知模块引导网络聚焦建筑特征并弱化背景噪声干扰;其次,采用细节对比增强机制从多个角度捕获鉴别性的空间细节特征;最后,引入逐级自适应特征融合策略由浅入深对各阶段特征进行逐步细化,完成跨层特征的有效融合.本方法在 ROxf 和 RPar 两大主流建筑数据集上的 mAP 指标分别达到了 80.84%(M),6 2.5 5%(H),90.04%(M)和 78.06%(H),有效提升了建筑图像检索的鲁棒性与准确性.
A multi-level adaptive fusion network was proposed for architectural image retrieval.It ad-dressed issues such as feature extraction difficulty,loss of detailed information,and insufficient feature fu-sion.First,asymmetric convolution blocks were introduced into the ResNet50 backbone.This improved fo-cus on the target architectural areas.Then,a global information awareness module was designed.It guided the network to focus on architectural features while suppressing background noise.Next,a detail contrast enhancement mechanism was employed.Discriminative spatial features were captured from multiple per-spectives.Finally,a layer-by-layer adaptive feature fusion strategy was adopted.It refined features progres-sively,achieving effective cross-layer fusion.The proposed method achieved mAP scores of 80.84%(M)and 62.55%(H)on the ROxf dataset.It also achieved 90.04%(M)and 78.06%(H)on the RPar data-set.Robustness and accuracy in architectural image retrieval were significantly improved.
张紫琴;肖青龙;范县成;汤巍;黄洪斌
安徽信息工程学院 电气与电子工程学院,安徽省芜湖市永和路1号 241000合肥华致能源科技有限公司,安徽省合肥市联东U谷33-2幢 231271安徽信息工程学院 电气与电子工程学院,安徽省芜湖市永和路1号 241000安徽信息工程学院 电气与电子工程学院,安徽省芜湖市永和路1号 241000安徽信息工程学院 电气与电子工程学院,安徽省芜湖市永和路1号 241000
信息技术与安全科学
建筑图像检索网络ResNet50全局信息感知特征融合细节对比增强
architectural imagesimage retrievalResNet50global information perceptionfeature fusiondetail contrast enhancement
《河北水利电力学院学报》 2026 (3)
63-72,10
安徽省高等学校自然科学研究计划项目(2023AH052918,2024AH050637)
评论