Remote Sensing for Natural Resources >
Small target detection in remote sensing images based on lightweight YOLOv7-tiny
Received date: 2024-03-15
Revised date: 2024-12-14
Online published: 2026-06-03
To address the issues of low detection accuracy caused by significant scale variations, complex scenes, and limited feature information of small targets in remote sensing images, as well as low detection efficiency resulting from the large parameter size and high complexity of current object detection models, this study proposes a lightweight YOLOv7-tiny model for remote sensing image detection. First, the network neck was improved by incorporating group shuffle convolution (GSConv) and VoV-GSCSP modules. This allows for sufficient detection accuracy while reducing computational costs and network complexity. Second, a dynamic head (DyHead) combined with an attention mechanism was adopted during prediction. The performance of the detection head was enhanced using multi-head self-attention across scale-aware feature layers, spatially-aware positions, and task-aware output channels. Finally, the loss function of the original model was optimized by integrating the normalized Wasserstein distance (NWD) metric for small-target assessment and a bounding box regression loss function based on the minimum point distance IoU (MPDIoU). This assists in enhancing robustness for small target detection. The experimental results demonstrate that the proposed algorithm achieved mAP@50 scores of 87.7% and 94.7% on the DIOR and RSOD datasets, respectively, indicating increases of 2.7 and 5.1 percentage points compared to the original YOLOv7-tiny model. Furthermore, the frames per second (FPS) increased by 12.2% and 11.9%, respectively. Therefore, the proposed algorithm can effectively enhance both the accuracy and real-time performance of small target detection from remote sensing images.
Key words: remote sensing images; object detection; YOLOv7-tiny; GSConv; MPDIoU; DyHead
XU Ziyao , YANG Wu , SHI Xiaolong . Small target detection in remote sensing images based on lightweight YOLOv7-tiny[J]. Remote Sensing for Natural Resources, 2025 , 37(4) : 1 -11 . DOI: 10.6046/zrzyyg.2024102
表1 DIOR及RSOD数据集信息Tab.1 Information about the DIOR dataset and the RSOD dataset |
| 属性 | DIOR | RSOD |
|---|---|---|
| 分类数/个 图像数/幅 实例数/个 年份 | 20 23 463 190 288 2019年 | 4 976 6950 2015年 |
表2 消融实验结果对比Tab.2 Ablation experiment Comparison of results |
| 序号 | NWD+ MPDIoU | GSConv+ VoV-GSCSP | Dyhead | mAP@ 0.5/% | 参数量/ 106个 |
|---|---|---|---|---|---|
| 1 | × | × | × | 85.0 | 6.1 |
| 2 | √ | × | × | 85.7 | 6.1 |
| 3 | × | √ | × | 86.2 | 5.6 |
| 4 | × | × | √ | 87.0 | 5.8 |
| 5 | √ | √ | × | 86.6 | 5.6 |
| 6 | √ | √ | √ | 87.7 | 5.4 |
表3 不同算法在DIOR数据集上的实验结果对比Tab.3 Experiment results comparison of different algorithms on the DIOR dataset |
| 方法 | mAP@0.5/% | 参数量/ 106个 | FPS/(帧·s-1) |
|---|---|---|---|
| Faster R-CNN | 75.8 | 28.5 | 17.4 |
| SSD | 64.1 | 27.1 | 66.1 |
| RetinaNet | 72.4 | 36.2 | 25.8 |
| YOLOv3 | 77.6 | 61.6 | 53.8 |
| YOLOv5s | 85.8 | 7.2 | 82.6 |
| YOLOv7 | 87.1 | 38.3 | 45.8 |
| YOLOv7-tiny | 85.0 | 6.1 | 76.8 |
| YOLOv8s | 86.6 | 11.1 | 86.1 |
| 本文方法 | 87.7 | 5.4 | 86.2 |
表4 不同算法在RSOD数据集上的实验结果对比Tab.4 Experiment results comparison of different algorithms on the RSOD dataset |
| 方法 | mAP@0.5/% | 参数量/ 106个 | FPS/(帧·s-1) |
|---|---|---|---|
| Faster R-CNN | 84.4 | 28.5 | 11.8 |
| SSD | 82.6 | 27.1 | 73.0 |
| RetinaNet | 86.5 | 36.2 | 22.4 |
| YOLOv3 | 86.1 | 61.6 | 50.9 |
| YOLOv5s | 90.6 | 7.2 | 79.3 |
| YOLOv7 | 94.2 | 38.3 | 42.7 |
| YOLOv7-tiny | 89.6 | 6.1 | 73.5 |
| YOLOv8s | 93.8 | 11.1 | 82.2 |
| 本文方法 | 94.7 | 5.4 | 82.3 |
图8-1 所提算法与YOLOv7-tiny在DIOR数据集上检测结果对比Fig.8-1 Comparison of detection results between the proposed algorithm and YOLOv7-tiny on the DIOR dataset |
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
付涵, 范湘涛, 严珍珍, 等. 基于深度学习的遥感图像目标检测技术研究进展[J]. 遥感技术与应用, 2022, 37(2):290-305.
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
张路青, 郭莹. 基于卷积神经网络的遥感图像目标检测识别[J]. 舰船电子工程, 2023, 43(5):49-53.
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
李安达, 吴瑞明, 李旭东. 改进YOLOv7的小目标检测算法研究[J]. 计算机工程与应用, 2024, 60(1):122-134.
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
/
| 〈 |
|
〉 |