GRACE-YOLOv7: An Improved Deep Learning Method for Object Detection in Remote Sensing Images
Year: 2026
Type: article
Abstract: The substantial variation in object scales, complex spatial distributions, and rich background information in remote sensing images can lead to inadequate feature representation, missed detections, and inaccurate localization of small and densely distributed objects. To address these issues, this study proposes GRACE-YOLOv7, an improved network for multi-scale object detection in remote sensing images based on YOLOv7. First, an ELAN-D module incorporating deformable convolution v3 (DCNv3) is constructed during deep feature extraction in the backbone. By adaptively adjusting the spatial sampling locations, the proposed module enhances the network’s ability to represent objects with different scales and geometric structures. Second, a Bi-level Routing Attention (BRA) mechanism is introduced at key feature interaction positions in the neck to strengthen information interaction among target-relevant regions and reduce the interference of complex background information with multi-scale feature fusion. On this basis, shallow high-resolution features are incorporated and an additional detection branch is introduced, extending the original three-scale detection structure to a four-scale detection structure to enhance the representation of small and densely distributed objects. Meanwhile, MPDIoU is adopted to replace the original CIoU as the bounding box regression loss to further improve object localization. Experimental results on the DOTA-v1.0 dataset show that GRACE-YOLOv7 achieves mAP0.5, mAP0.75, Recall, Precision, and F1 scores of 86.09%, 65.42%, 84.40%, 82.60%, and 0.8347, respectively. Compared with YOLOv7, mAP0.5 and mAP0.75 increase by 2.35 and 3.28 percentage points, respectively, while Recall and Precision increase by 2.81 and 1.63 percentage points, respectively, and F1 increases from 0.8093 to 0.8347. These results demonstrate that GRACE-YOLOv7 improves multi-scale object detection performance on the DOTA-v1.0 dataset, particularly for small and densely distributed objects.
Source: Sensors
DOI: 10.3390/s26217092
Authors: Xiang Gao, Changbao Yang, Liguo Han
Institution: Jilin University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7225661012
FWCI:
Cites:
Cited by:
Related to: