遥感
计算机科学
目标检测
转化(遗传学)
比例(比率)
人工智能
计算机视觉
特征(语言学)
融合
图像融合
传感器融合
变更检测
对象(语法)
图像增强
特征提取
遥感应用
模式识别(心理学)
边缘检测
激光雷达
作者
Guoao Zhou,Bingrong Xu,Yixin Su,Cheng Lian,Zhigang Zeng
标识
DOI:10.1109/tgrs.2026.3748977
摘要
Remote sensing object detection generally faces challenges such as large scale variations and severe background interference. Existing methods mainly improve detection performance by enlarging the receptive field or enhancing contextual dependencies. However, the former tends to weaken the texture information of small objects, while the latter is often constrained by limited local context, resulting in insufficient feature representation. To address these issues, we propose a Multi-kernel Scale Transformation Enhancement Network(MSTENet). Specifically, Scale Transformation Module reconstructs multi-scale receptive fields to capture objects of varying scales while modeling long-range contextual information. To alleviate the loss of small-object features caused by multi-scale receptive field modeling, Fine-grained Enhancement Module is cascaded after STM to compensate for the fine-grained information of small objects that may be lost during scale transformation while enhancing local contextual representation. Furthermore, to balance the collaborative effects of STM and FEM, we introduce a Gated Feature Fusion mechanism that adaptively integrates the global features from STM with the local features from FEM, thereby enhancing foreground–background discrimination and category-boundary representation. Experimental results show that MSTENet achieves mAP scores of 82.72%, 90.67%, and 68.84% on the DOTA-v1.0, HRSC2016, and DIOR datasets, respectively, demonstrating the effectiveness of the proposed method.
科研通智能强力驱动
Strongly Powered by AbleSci AI