<p>Small object detection in remote sensing imagery remains a challenging problem in computer vision. To address the inherent limitations of aerial imagery, such as densely packed objects with insufficient detail and occlusions caused by complex backgrounds, this study proposes LKR-DETR, an innovative object detection in remote sensing imagery framework based on RT-DETR. We propose a lightweight and efficient feature extraction module with large kernel convolution, which expands the receptive field while reducing parameters and computational costs. Furthermore, we present a novel multi-scale feature fusion structure based on wavelet transform convolution that effectively utilizes low-frequency information from low-level feature maps. Additionally, we introduce a lightweight image restoration module utilizing large kernel convolutions, which effectively recovers previously undetected details of small objects. To improve bounding box regression accuracy, the original GIoU loss is replaced with a Focaler-DIoU loss function. Compared to the benchmark model RT-DETR, the LKR-DETR model achieves a 2.5% improvement in <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1622_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="55" /> </InlineMediaObject> <EquationSource Format="TEX">$mAP_{0.5}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>m</mi> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mn>0.5</mn> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> and a 2.0% improvement in <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1622_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="79" /> </InlineMediaObject> <EquationSource Format="TEX">$mAP_{0.5:0.95}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>m</mi> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mn>0.5</mn> <mo>:</mo> <mn>0.95</mn> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> on the VisDrone2019-DET dataset, a 1.7% and 4.4% improvement on the DOTAv1.5 dataset, and a 3.4% and 2.4% improvement on the HIT-UAV dataset, while also reducing the parameter count and model size. Relative to other cutting-edge models, LKR-DETR attains superior detection accuracy while maintaining relatively low computational complexity, establishing it as an efficient solution for small object detection in remote sensing imagery.</p>

LKR-DETR: small object detection in remote sensing images based on multi-large kernel convolution

期刊论文
发表日期：

2025年1月17日

查看全文

Ying Dong,
Fucheng Xu,
Jiahao Guo

摘要

Small object detection in remote sensing imagery remains a challenging problem in computer vision. To address the inherent limitations of aerial imagery, such as densely packed objects with insufficient detail and occlusions caused by complex backgrounds, this study proposes LKR-DETR, an innovative object detection in remote sensing imagery framework based on RT-DETR. We propose a lightweight and efficient feature extraction module with large kernel convolution, which expands the receptive field while reducing parameters and computational costs. Furthermore, we present a novel multi-scale feature fusion structure based on wavelet transform convolution that effectively utilizes low-frequency information from low-level feature maps. Additionally, we introduce a lightweight image restoration module utilizing large kernel convolutions, which effectively recovers previously undetected details of small objects. To improve bounding box regression accuracy, the original GIoU loss is replaced with a Focaler-DIoU loss function. Compared to the benchmark model RT-DETR, the LKR-DETR model achieves a 2.5% improvement in $mAP_{0.5}$ $m A P_{0.5}$ and a 2.0% improvement in $mAP_{0.5:0.95}$ $m A P_{0.5 : 0.95}$ on the VisDrone2019-DET dataset, a 1.7% and 4.4% improvement on the DOTAv1.5 dataset, and a 3.4% and 2.4% improvement on the HIT-UAV dataset, while also reducing the parameter count and model size. Relative to other cutting-edge models, LKR-DETR attains superior detection accuracy while maintaining relatively low computational complexity, establishing it as an efficient solution for small object detection in remote sensing imagery.