东北石油大学学报2026，Vol.50Issue(2)：109-122,14.DOI:10.3969/j.issn.2095-4107.2026.02.008
基于显著性先验和金字塔转换机制的Trans-Sal Det小目标油井检测模型
Trans-Sal Det small object detection model based on saliency prior and pyramid transformation mechanism
摘要
Abstract
To address the difficulty of accurately detecting small oil well targets in oilfield remote sensing images under complex surface environments and background interference,this paper proposes a Trans-Sal Det small object detection model that integrates saliency priors with a pyramid transformation mech-anism.A saliency generation module based on Tiny-U-Net is constructed to produce high-resolution sa-liency maps,highlighting potential oil well regions while suppressing redundant background informa-tion.The saliency maps are concatenated with the original remote sensing images at the channel level and fed into a pyramid Transformer encoder.Through a multi-scale window self-attention mechanism,cross-layer feature modeling is achieved,effectively integrating low-level detailed features with high-lev-el semantic information.A cross-scale feature fusion strategy is introduced to enhance the model's rep-resentation capability for oil well targets of different scales.Comparative and ablation experiments are conducted on typical oilfield remote sensing datasets to systematically validate the model performance.The results demonstrate that Trans-Sal Det exhibits superior small object detection performance in com-plex backgrounds,occlusion,and low-contrast scenarios,with significantly improved recall and detec-tion accuracy compared to mainstream methods.The introduction of saliency priors effectively guides at-tention to focus on key regions,enhancing the Transformer's perception of small targets.The proposed method provides an efficient and feasible technical approach for automated oilfield remote sensing detec-tion and intelligent monitoring,with strong potential for practical applications.关键词
显著性先验/金字塔转换机制/交叉尺度融合/油井检测/动态显著性/Transformer/小目标检测Key words
saliency prior/pyramid transformation mechanism/cross-scale feature fusion/oil well de-tection/dynamic saliency/Transformer/small object detection分类
能源科技引用本文复制引用
赵梓翔,李佳慧,步贤业,穆树娟,隋杨..基于显著性先验和金字塔转换机制的Trans-Sal Det小目标油井检测模型[J].东北石油大学学报,2026,50(2):109-122,14.基金项目
国家自然科学基金面上项目(62573112) （62573112）
黑龙江省自然科学基金优秀青年基金项目(YQ2023F003) （YQ2023F003）
黑龙江省博士后特别资助项目(LBH-TZ2505) （LBH-TZ2505）
