由于受到硬件条件的限制,通常难以获得具有高分辨率(HR)的遥感图像。利用单幅图像超分辨率(SISR)技术对低分辨率(LR)遥感图像进行超分辨率重建是获取高分辨率遥感图像的常用方法。近年来,在图像超分辨率领域引入的卷积神经网络(CNN)改进了图像重建性能。然而,现有的基于CNN的超分辨率模型通常使用低阶注意力机制提取深层特征,其表征能力有待提高,且常规卷积的感受野有限,缺乏对远距离依赖关系的学习。为了解决以上问题,提出了一种基于递归门控卷积的遥感图像超分辨率方法RGCSR。该方法引入递归门控卷积g n Conv学习全局依赖和局部细节,通过高阶空间交互来获取高阶特征。首先,使用由高阶交互子模块(HorBlock)和前馈神经网络(FFN)组成的高阶交互——前馈神经网络模块(HFB)提取高阶特征。其次,利用由通道注意力(CA)和g n Conv构建的特征优化模块(FOB)优化各个中间模块的输出特征。最后,在多个数据集上的对比结果表明,RGCSR比现有的基于CNN的超分辨率方法具备更好的重建性能和视觉效果。
无人机航拍图像目标检测在民用和军事领域具有重要的应用价值。针对无人机航拍图像中目标小、尺度变化大和背景干扰等因素导致检测精度低、定位不准确的问题,提出一种改进YOLOv10n的无人机航拍图像目标检测算法。首先将C2f模块进行改进,利用递归门控卷积(gnConv)与c2f融合二次创新得到C2f-GConv模块,以适应航拍图像中物体的形变和尺度变化。同时将骨干网络替换成Efficientformerv2,使得EfficientFormerV2在保持类似MobileNetV2大小和速度的同时,比MobileNetV2高约4%的top-1精度,明显提高了模型的效率和性能。在VisDrone2019数据集上进行对比实验和消融实验,mAP50值较基线模型提升了3.2%,检测速度FPS达到90帧/s,能够满足实时性的检测需求。与主流算法进行对比实验,所提算法表现优于目前主流算法。Aerial target detection in unmanned aerial vehicle (UAV) imagery holds significant application value in both civilian and military fields. To address the challenges of low detection accuracy and imprecise localization caused by small targets, large scale variations, and background interference in UAV imagery, an improved YOLOv10n algorithm for aerial image target detection is proposed. Firstly, the C2f module is enhanced by integrating the recursive gated convolution (gnConv) with the C2f for a second innovation, resulting in C2f-GConv adapting to the deformation and scale changes of objects in aerial images. Meanwhile, the backbone network is replaced with EfficientFormerV2, which maintains size and speed similar to MobileNetV2 but achieves about 4% higher top-1 accuracy than MobileNetV2, significantly improving the model’s efficiency and performance. Comparative and ablation experiments are conducted on the VisDrone2019 dataset, with the mAP50 value increasing by 3.2% over the baseline model and a detection speed of FPS reaching 90 frames per second, meeting the real-time detection requirements. Comparative experiments with mainstream algorithms show that the proposed algorithm outperforms current mainstream algorithms.
暂无评论