E-mail Alert
RSS
-
摘要:
针对图像中背景复杂、目标小、分布密集等问题，提出了一种改进的DES-YOLO方法。通过引入可变形注意力模块(DAM)，网络可动态关注关键区域，提高物体识别和定位精度；采用高效交并比(EIoU)损失函数，减少低质量样本影响，增强泛化能力和检测精度；在网络头部加入一层160 pixel×160 pixel的浅层特征图，加强小目标特征提取；并使用分步训练策略提升模型性能。实验结果表明，该模型在遥感数据集上的mAP@50提升了1.4%，在纺织数据集上提升了1.7%，验证了DES-YOLO的广泛适用性与有效性。
Abstract:To address the challenges of complex backgrounds, small targets, and dense distributions in images, an improved method called DES-YOLO is proposed. By introducing the deformable attention module (DAM), the network can dynamically focus on key regions, improving object recognition and localization accuracy. The efficient intersection over union (EIoU) loss function is employed to reduce the impact of low-quality samples, enhancing the model's generalization ability and detection accuracy. A shallow feature map layer of 160 pixel×160 pixel is added to the network head to strengthen small target feature extraction. A stepwise training strategy is also adopted to further improve model performance. Experimental results show that the mAP@50 of the model increased by 1.4% on the remote sensing dataset and by 1.7% on the textile dataset, demonstrating the broad applicability and effectiveness of DES-YOLO.
-
Key words:
- object detection /
- deformable attention /
- EIoU /
- shallow features /
- stepwise training strategy /
- DES-YOLO
-
Overview: In image analysis, detecting objects accurately remains a significant challenge due to the complexity of backgrounds, the small size of targets, and their dense distribution. To address these issues, we propose an advanced detection method named DES-YOLO. This method incorporates several innovative techniques to enhance the performance of object detection in remote sensing imagery. Firstly, we introduce a deformable attention module (DAM), which allows the network to dynamically adjust its focus on crucial areas of the image. This module enables the network to better recognize and localize objects by concentrating on significant regions and ignoring irrelevant background noise. Secondly, we implement the efficient intersection over union (EIoU) loss function, designed to mitigate the influence of low-quality samples. This loss function improves the generalization ability and detection accuracy of the model, ensuring more precise object localization. Furthermore, we augment the network head with an additional shallow feature map layer of 160 pixel×160 pixel. This enhancement specifically targets extracting features from small objects, often challenging to detect in remote-sensing images. By capturing more detailed information, this layer significantly boosts the detection capability for small-sized targets. Additionally, we employ a stepwise training strategy to refine the model's performance progressively. This training approach helps stabilise the learning process and improves the robustness of the model, leading to superior detection outcomes. Our experimental results are compelling. The improved DES-YOLO model demonstrates a 1.4% increase in the mean average precision (mAP@0.5) on a standard remote sensing dataset. To further validate the model's effectiveness, we conducted extended experiments on a textile dataset, where the model achieved an impressive mAP@0.5 increase of 1.7%. These results not only highlight the improvements brought by our method but also confirm its versatility and applicability to various types of datasets. In conclusion, DES-YOLO represents a significant advancement in object detection, offering enhanced accuracy and reliability. Integrating the deformable attention module, EIoU loss function, shallow feature enhancement, and stepwise training collectively contribute to its superior performance. Our research demonstrates the potential of DES-YOLO to set a new benchmark in object detection, paving the way for future developments and applications.
-
-
图 6 不同网络模型检测精度对比
Figure 6. Comparison of detection accuracy of different network models
表 1 实验环境与配置
Table 1. Experimental environment and configuration
Type Configuration GPU NVIDIA GeFore RTX4090 CPU 13th Gen Intel(R) Core(TM) i7-13620H CUDA 11.7 Deep learning framework Pytorch Python 3.12 
下载: 导出CSV
表 2 遥感目标检测数据集
Table 2. Remote sensing target detection data set
Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv5s 0.937 0.899 0.929 0.562 YOLOv5s+CBAM 0.939 0.884 0.923 0.511 YOLOv5s+CA 0.922 0.886 0.931 0.551 YOLOv5s+DA 0.945 0.898 0.942 0.561
下载: 导出CSV
表 3 纺织物瑕疵检测数据集
Table 3. Textile defect detection data set
Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv5s 0.350 0.322 0.276 0.118 YOLOv5s+CBAM 0.382 0.290 0.282 0.141 YOLOv5s+CA 0.237 0.296 0.219 0.086 YOLOv5s+DA 0.350 0.342 0.285 0.121
下载: 导出CSV
表 4 遥感目标探测损失函数的效果比较
Table 4. Effect comparison of remote sensing target detection loss function
Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv5s+DA+CIoU 0.945 0.898 0.942 0.561 YOLOv5s+DA+EIoU 0.936 0.914 0.944 0.573 YOLOv5s+DA+SIoU 0.933 0.922 0.935 0.571 YOLOv5s+DA+WIoU 0.848 0.842 0.884 0.506
下载: 导出CSV
表 5 纺织品缺陷检测损失函数的效果比较
Table 5. Effect comparison of textile defect detection loss function
Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv5s+DA+CIoU 0.350 0.342 0.285 0.121 YOLOv5s+DA+EIoU 0.392 0.315 0.281 0.130 YOLOv5s+DA+SIoU 0.382 0.272 0.283 0.144 YOLOv5s+DA+WIoU 0.357 0.292 0.258 0.103
下载: 导出CSV
表 6 遥感目标检测
Table 6. Remote sensing target detection
Params/M Precision Recall mAP@0.5 mAP@0.5:0.95 Ship YOLOv5s+DA 8.1 0.936 0.914 0.944 0.573 0.973 YOLOv5s+DA+STD 8.7 0.963 0.903 0.943 0.613 0.98
下载: 导出CSV
表 7 纺织物瑕疵检测
Table 7. Textile defect detection
Params/M Precision Recall mAP@0.5 mAP@0.5:0.95 Knot head YOLOv5s+DA 8.1 0.392 0.315 0.281 0.13 0.303 YOLOv5s+DA+STD 8.7 0.454 0.258 0.293 0.14 0.311
下载: 导出CSV
表 8 消融实验结果
Table 8. Results of ablation experiments
Params/M Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv5s 7.0 0.937 0.899 0.929 0.562 YOLOv5s+DA 8.1 0.945 0.898 0.942 0.561 YOLOv5s+DA+EIoU 8.1 0.936 0.914 0.944 0.573 YOLOv5s+DA+EIoU+STD 8.7 0.963 0.903 0.943 0.613
下载: 导出CSV
表 9 对比实验结果
Table 9. Results of ablation experiments
Params/M Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv5s 7.1 0.350 0.322 0.276 0.118 YOLOv5s+DA 8.1 0.350 0.342 0.285 0.121 YOLOv5s+DA+EIoU 8.1 0.392 0.315 0.281 0.130 YOLOv5s+DA+EIoU+STD 8.7 0.454 0.258 0.293 0.140
下载: 导出CSV
表 10 对比实验结果
Table 10. Results of comparison experiments
Params/M Precision Recall mAP@0.5 mAP@0.5:0.95 GFLOPs/G Faster-RCNN 41.1 0.861 0.957 0.901 0.559 33.2 YOLOv3-tiny 8.6 0.952 0.860 0.929 0.544 12.9 YOLOv3 61.5 0.956 0.918 0.952 0.602 155.4 YOLOv5s 7.0 0.937 0.899 0.929 0.562 15.8 YOLOv5m 20.9 0.864 0.832 0.888 0.523 48.0 YOLOv7-tiny 6.0 0.786 0.631 0.768 0.386 13.3 YOLOv8s 11.1 0.906 0.867 0.932 0.601 28.5 DES-YOLO 8.7 0.963 0.903 0.943 0.613 27.9
下载: 导出CSV
-
[1] 张阳婷, 黄德启, 王东伟, 等. 基于深度学习的目标检测算法研究与应用综述[J]. 计算机工程与应用, 2023, 59(18): 1−13. doi: 10.3778/j.issn.1002-8331.2305-0310
Zhang Y T, Huang D Q, Wang D W, et al. Review on research and application of deep learning-based target detection algorithms[J]. Comput Eng Appl, 2023, 59(18): 1−13. doi: 10.3778/j.issn.1002-8331.2305-0310
[2] 童康, 吴一全. 基于深度学习的小目标检测基准研究进展[J]. 电子学报, 2024, 52(3): 1016−1040. doi: 10.12263/DZXB.20230624
Tong K, Wu Y Q. Research advances on deep learning based small object detection benchmarks[J]. Acta Electron Sin, 2024, 52(3): 1016−1040. doi: 10.12263/DZXB.20230624
[3] Jiang T, Mu X D, Wei X, et al. Research progress of single-stage small target detection based on deep learning[C]//2022 4th International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM), Hamburg, Germany, 2022: 893–898. https://doi.org/10.1109/AIAM57466.2022.00180.
[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. doi: 10.11873/j.issn.1004-0323.2022.2.0290
Fu H, Fan X T, Yan Z Z, et al. Progress of object detection in remote sensing images based on deep learning[J]. Remote Sens Technol Appl, 2022, 37(2): 290−305. doi: 10.11873/j.issn.1004-0323.2022.2.0290
[5] Wang X L, Ban Y, Guo H M, et al. Deep learning model for target detection in remote sensing images fusing multilevel features[C]//IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 2019: 250–253. https://doi.org/10.1109/IGARSS.2019.8898759.
[6] Dong R C, Xu D Z, Zhao J, et al. Sig-NMS-based faster R-CNN combining transfer learning for small target detection in VHR optical remote sensing imagery[J]. IEEE Trans Geosci Remote Sens, 2019, 57(11): 8534−8545. doi: 10.1109/TGRS.2019.2921396
[7] Mehta A, Jain R. An analysis of fabric defect detection techniques for textile industry quality control[C]//2023 World Conference on Communication & Computing (WCONF), Raipur, India, 2023: 1–5. https://doi.org/10.1109/WCONF58270.2023.10235154.
[8] Karlekar V V, Biradar M S, Bhangale K B. Fabric defect detection using wavelet filter[C]//2015 International Conference on Computing Communication Control and Automation, Pune, India, 2015: 712–715. https://doi.org/10.1109/ICCUBEA.2015.145.
[9] Alimohamadi H, Ahmadyfard A, Shojaee E. Defect detection in textiles using morphological analysis of optimal Gabor wavelet filter response[C]//2009 International Conference on Computer and Automation Engineering, Bangkok, Thailand, 2009: 26–30. https://doi.org/10.1109/ICCAE.2009.43.
[10] 程汉权, 熊继平, 陈经纬. 布匹瑕疵检测算法研究进展[J]. 计算机时代, 2023, (11): 16−21. doi: 10.16644/j.cnki.cn33-1094/tp.2023.11.004
Cheng H Q, Xiong J P, Chen J W. Research progress of fabric defect detection[J]. Comput Era, 2023, (11): 16−21. doi: 10.16644/j.cnki.cn33-1094/tp.2023.11.004
[11] Zhou H, Jang B, Chen Y X, et al. Exploring faster RCNN for fabric defect detection[C]//2020 Third International Conference on Artificial Intelligence for Industries (AI4I), Irvine, CA, USA, 2020: 52–55. https://doi.org/10.1109/AI4I49448.2020.00018.
[12] Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection, [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017: 936-944. https://doi.org/10.1109/CVPR.2017.106
[13] Li J F, Zhu Y W, Chen M X, et al. Research on underwater small target detection algorithm based on improved YOLOv3[C]//2022 16th IEEE International Conference on Signal Processing (ICSP), Beijing, China, 2022: 76–80. https://doi.org/10.1109/ICSP56322.2022.9965317.
[14] Cao K Y, Cui X, Piao J C. Smaller target detection algorithms based on YOLOv5 in safety helmet wearing detection[C]//2022 4th International Conference on Robotics and Computer Vision (ICRCV), Wuhan, China, 2022: 154–158. https://doi.org/10.1109/ICRCV55858.2022.9953233.
[15] 张冲, 黄影平, 郭志阳, 等. 基于语义分割的实时车道线检测方法[J]. 光电工程, 2022, 49(5): 210378. doi: 10.12086/oee.2022.210378
Zhang C, Huang Y P, Guo Z Y, et al. Real-time lane detection method based on semantic segmentation[J]. Opto-Electron Eng, 2022, 49(5): 210378. doi: 10.12086/oee.2022.210378
[16] Luo M M, Huang J H, Sun X Y, et al. Small target forest fire recognition method based on deep learning[C]//2023 IEEE 3rd International Conference on Information Technology, Big Data and Artificial Intelligence (ICIBA), Chongqing, China, 2023: 593–597. https://doi.org/10.1109/ICIBA56860.2023.10165608.
[17] Li R Z, Chen Y J, Sun C Y, et al. Improved algorithm for small target detection of traffic signs on YOLOv5s[C]//2023 4th International Conference on Intelligent Computing and Human-Computer Interaction (ICHCI), Guangzhou, China, 2023: 339–344. https://doi.org/10.1109/ICHCI58871.2023.10278065.
[18] 陈旭, 彭冬亮, 谷雨. 基于改进YOLOv5s的无人机图像实时目标检测[J]. 光电工程, 2022, 49(3): 210372. doi: 10.12086/oee.2022.210372
Chen X, Peng D L, Gu Y. Real-time object detection for UAV images based on improved YOLOv5s[J]. Opto-Electron Eng, 2022, 49(3): 210372. doi: 10.12086/oee.2022.210372
[19] Ge R, Mao Y L, Li S, et al. Research on ship small target detection in SAR image based on improved YOLO-v7[C]//2023 International Applied Computational Electromagnetics Society Symposium (ACES-China), Hangzhou, China, 2023: 1–3. https://doi.org/10.23919/ACES-China60289.2023.10249265.
[20] Chen Z G, Liu G X, Fan S W. Research on target detection algorithm based on improved YOLO[C]//2022 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML), Xi’an, China, 2022: 485–489. https://doi.org/10.1109/ICICML57342.2022.10009683.
[21] Zhang H Y, Deng L X, Bi L Y, et al. Small object detection algorithm based on improved yolov5[C]//2023 IEEE International Conference on Control, Electronics and Computer Technology (ICCECT), Jilin, China, 2023: 280–283. https://doi.org/10.1109/ICCECT57938.2023.10141436.
[22] Pu J T, Zhang H Y, Yuan M D, et al. ACN-YOLO: an algorithm for small target detection in aerial images[C]//2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), Haikou, China, 2023: 158–163. https://doi.org/10.1109/PRAI59366.2023.10331968.
[23] Lin T. Focal loss for dense object detection[Z]. arXiv:1708.02002, 2017. https://doi.org/10.48550/arXiv.1708.02002
[24] Wang F L, Su J Y. Based on the improved YOLOV3 small target detection algorithm[C]//2021 IEEE 4th Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), Chongqing, China, 2021: 2155–2159. https://doi.org/10.1109/IMCEC51613.2021.9482076.
[25] 栾庆磊, 常昕昱, 吴叶, 等. PAW-YOLOv7: 河道微小漂浮物检测算法[J]. 光电工程, 2024, 51(4): 240025. doi: 10.12086/oee.2024.240025
Luan Q L, Chang X Y, Wu Y, et al. PAW-YOLOv7: algorithm for detection of tiny floating objects in river channels[J]. Opto-Electron Eng, 2024, 51(4): 240025. doi: 10.12086/oee.2024.240025
[26] Xia Z F, Pan X R, Song S J, et al. Vision transformer with deformable attention[C]//Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, 2022: 4784–4793. https://doi.org/10.1109/CVPR52688.2022.00475.
[27] Cheng G, Han J W, Zhou P C, et al. Multi-class geospatial object detection and geographic image classification based on collection of part detectors[J]. ISPRS J Photogramm Remote Sens, 2014, 98: 119−132. doi: 10.1016/j.isprsjprs.2014.10.002
[28] Cheng G, Han J W. A survey on object detection in optical remote sensing images[J]. ISPRS J Photogramm Remote Sens, 2016, 117: 11−28. doi: 10.1016/j.isprsjprs.2016.03.014
[29] Cheng G, Zhou P C, Han J W. Learning rotation-invariant convolutional neural networks for object detection in VHR optical remote sensing images[J]. IEEE Trans Geosci Remote Sens, 2016, 54(12): 7405−7415. doi: 10.1109/TGRS.2016.2601622
[30] 天池. 布匹瑕疵检测数据集[EB/OL]. 2020[2024-12-10].https://tianchi.aliyun.com/dataset/dataDetail?dataId=79336.
Tianchi. Smart diagnosis of cloth flaw dataset[EB/OL]. 2020[2024-12-10]. https://tianchi.aliyun.com/dataset/dataDetail?dataId=79336.
-
点击扫一扫
图(8)
表(10)
计量
- 文章访问数:
- PDF下载数:
- 施引文献: 0
