伪装目标智能检测旨在利用有限特征信息，智能识别隐藏在复杂环境中的目标。现有伪装目标智能检测方法存在低显著性特征、类间差异小、数据集标注成本高、小像素点特征提取手段有限及伪装场景下目标关键特征提取性能不佳等问题。针对以上问题，本文提出了基于空天无人平台辅助信息的伪装目标智能检测新方法。首先，提出了一种改进的深度卷积生成对抗网络，通过设计新的损失函数和注意力机制模块提高图像生成质量，从而扩充数据集；然后，设计了一种改进型的Yolov11网络——Yolov11-Codattention，该网络的独特之处在于利用空天无人平台辅助信息，内嵌了上下文信息和感受野机制两类模块，以解决低显著性特征和类间差异小等问题。通过两类模块的引入，所设计的网络在目标特征提取能力大幅增强的同时，对模型参数量的需求大幅减小。基于公开的空天迷彩数据集，进行了伪装目标检测仿真实验，实验结果表明：提出的Yolov11-Codattention算法与传统Yolov11算法相比，在召回率和mAP@50性能指标上分别提高了7.0%和4.8%，且实时性达到40FPS；通过与七种常用目标检测算法的对比，发现Yolov11-Codattention具有更高的伪装目标检测精度。外场嵌入式部署实验结果表明：Yolov11-Codattention的平均置信度达到0.63，实时性达到30FPS，符合工程应用需求。实验结果充分验证了所设计的Yolov11-Codattention算法在伪装目标检测任务中的有效性。
Camouflage target detection aims to identify targets hidden in complex environments using limited feature information. Existing camouflage target detection methods have problems such as low saliency features, small inter-class differences, high annotation costs for datasets, limited means of extracting small pixel features, and poor performance in extracting key features of targets in camouflage scenarios. To address these issues, this paper proposes a new camouflage target detection method based on an improved Yolov11 network. Firstly, an improved deep convolutional generative adversarial network is proposed to expand the dataset, and enhance the image generation quality by designing a new loss function and embedding an attention mechanism module. Then, an improved Yolov11 network - Yolov11-Codattention is designed. The uniqueness of this network lies in the embedding of a newly designed module based on context information and receptive field mechanism to solve problems such as low saliency features and small inter-class differences. By introducing these two types of modules, the designed network significantly enhances the target feature extraction capability while reducing the demand for model parameters. Based on the public camouflage dataset, camouflage target detection simulation experiments were conducted. The experimental results show that the proposed Yolov11-Codattention algorithm improves the recall rate and mAP@50 performance indicators by 7.0% and 4.8% respectively compared with the traditional Yolov11 algorithm, and the real-time performance reaches 40FPS. Through comparison with seven commonly used target detection algorithms, it is found that Yolov11-Codattention has higher camouflage target detection accuracy. The results of the field embedded deployment experiments show that the average confidence of Yolov11-Codattention reaches 0.63, and the real-time performance reaches 30FPS, meeting the engineering application requirements. The experimental results fully verify the effectiveness of the designed Yolov11-Codattention algorithm in the camouflage target detection task.
[1] Galun M, Sharon E, Basri R, and Brandt A, Texture seg-mentation by multiscale aggregation of filter responses and shape elements[C].Proceedings Ninth IEEE Interna-tional Conference on Computer Vision, Nice, France, 2003: 716- 718.
[2]Boult T E, Micheals R J, Xiang G, et al.Into the woods: visual surveillance of noncooperative and camouflaged targets in complex outdoor settings[J].Proceedings of the IEEE, 2001, 89(10):1382-1402
[3]Boot W R, Neider M B, Kramer A F.Training and transfer of training in the search for camouflaged targets[J].At-tention, Perception, & Psychophysics, 2009, 71(4):950-963
[4]Zhang X, Zhu C, Wang S, et al.A Bayesian Approach to Camouflaged Moving Object Detection[J].IEEE Trans-actions on Circuits and Systems for Video Technology, 2017, 27(9):2001-2013
[5] Galun, Sharon, Basri, et al.Texture segmentation by mul-tiscale aggregation of filter responses and shape elements [C]. Proceedings of the Ninth IEEE International Confer-ence on Computer Vision (ICCV), Nice, France, 2003: 716-723.
[6]Neider M B, Zelinsky G J.Searching for camouflaged targets: Effects of target background similarity on visual search[J].Vision Research, 2006, 46(14):2217-2235
[7] Bhajantri N U, Nagabhushan P.Camouflage Defect Identi-fication: A Novel Approach [C]. Proceedings of the 9th International Conference on Information Technology (ICIT' 06), India, 2006:145-148.
[8]武国晶，吕绪良，邢海宁，等.三维凸面分析法在迷彩伪装检测中的应用[J].解放军理工大学学报：自然科学版, 2015, 16(6):582-586
[9]Tankus A, Yeshurun Y.Convexity-Based Visual Camou-flage Breaking[J].Computer Vision and Image Under-standing, 2001, 82(3):208-237
[10]Fan D-P, Ji G-P, Sun G, Cheng M-M, Shen J, and Shao L.Camouflaged object detection[J].in Proc. Comput. Vis. Pattern Recognit. (CVPR), 2020, 44(10):6024-6042
[11]Le T N, Nguyen T V, Nie Z, et al.Anabranch network for camouflaged object segmentation[J].Computer Vision and Image Understanding, 2019, 184(无):45-56
[12] Mei H, Ji G-P, Wei Z, Yang X, Wei X, and Fan D-P, Camouflaged object segmentation with distraction min-ing[C].in Proc. Comput. Vis. Pattern Recognit. (CVPR), USA, 2021: 8772–8781.
[13] Carion N, Massa F, Synnaeve G, Usunier N, Kirillov A, Zagoruyko S.End-to-End Object Detection with Trans-formers[C]. Computer Vision-ECCV 2020，Springer, Cham, 2020:213-229.
[14]刘文婷, 卢新明.基于计算机视觉的研究进展[J].计算机工程与应用, 2022, 58(06):1-16
[15]李科岑, 王晓强, 林浩, 等.深度学习中的单阶段小目标检测方法综述[J].计算机科学与探索, 2022, 16(1):41-58
[16]邵延华, 张铎, 楚红雨, 等.基于深度学习的目标检测综述[J].电子与信息学报, 2022, 44(10):3697-3708
[17] Redmon J, Farhadi A, YOLO9000: Better, Faster, Stronger[C].2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017: 6517-6525.
[18]Redmon J, Farhadi A, YOLOv3: An incremental im-provement[J]. Computer Science, 2..[J].rXiv: 1804.02767, [J]., :-
[19]Li H, Li J, Wei H, et al.Slim-neck by GSConv: A light-weight-design for real-time detector architec-tures[J].Computer Vision and Pattern Recognition, 2024, arXiv.2206.02424(无):1-11
[20] Varghese R, Sambath M, YOLOv8: A Novel Ob-ject Detection Algorithm with Enhanced Performance and Robustness[C].2024 International Conference on Ad-vances in Data Engineering and Intelligent Computing Systems, India, 2024:1-14.
[21] Woo S, Park J, Lee J-Y, et al.CBAM: Convolutional Block Attention Module [C]. Proceedings of the Euro-pean Conference on Computer Vision (ECCV), Springer, Cham.2018: 3-19.
CJA