Based on the YOLOv4 algorithm, this paper optimizes the number and size of the a priori anchor frames by using K-means re-clustering to reduce the computation of the network; introduces a dynamic activation function to adapt to the input features of the network and improve the generalization ability of the model; improves the PANet structure and extends the shallow features of feature fusion. The experimental results show that the improved YOLOv4 algorithm has improved the accuracy and detection speed for the recognition of densely arranged target images in remote sensing images, which meeting the requirements for small target detection in remote sensing images.
早期的目标检测模型是由人工制作的特征提取器组成,如Viola-Jones检测器[1]、方向梯度直方图(Histogram of Oriented Gradient,HOG)[2]等,这些模型速度慢、准确率低,应用环境改变后效果不佳。卷积神经网络(CNNs)的提出改变了视觉感知。AlexNet在2012年ImageNet大规模视觉识别挑战赛(ILSVRC)中的应用[3],激发了CNNs在计算机视觉中应用的进一步研究。如今,目标检测应用在自动驾驶、人脸识别等多领域[4]。
VIOLAP, JONESM.Rapid object detection using a boosted cascade of simple features[C]//Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,2001:511-518.
[2]
DALALN, TRIGGSB.Histograms of oriented gradients for human detection[C]//IEEE Computer Society Conference on Computer Vision and Pattern Recognition(CVPR’05),2005:1063-6919.
[3]
KRIZHEVSKYA, SUTSKEVERI, HINTONG.Image Net classification with deep convolutional neural networks[J].Advances in Neural Information Processing Systems,2012,25(2): 84-90.
HOUYueqian, ZHANGLihong. Multi-scale object detection based on Transformer[J]. Journal of Test and Measurement Technology, 2023, 37(4): 342-347. (in Chinese)
GAOQian, PANYang, ZHULei,et al.An improved YOLOv5 remote sensing target detection method based on SIOU function[J].Changjiang Information & Communication,2022,35(11):5-8. (in Chinese)
[8]
BOCHKOVSKIYA, WANGC Y, LIAOH.YOLOv4:Optimal speed and accuracy of object detection[J]. arXiv.org.2020.DOI:10.48550/arXiv.2004.10934 .
[9]
LIUC, FENGL, LIUG, WANGH B,et al.Bottom-up broadcast neural network for music genre classification[J].Multimedia Tools and Applications,2020(1):1-19.
[10]
GORBANA N, MIRKESE M, TUKINI Y.How deep should be the depth of convolutional neural networks:abackyard dog case study[J].Cognitive Computation,2020,12(1):388-397.
[11]
YIT.Exponential stability of pseudo almost periodic solutions for fuzzy cellular neural networks with time-varying delays[J].Neural Processing Letters,2019: 49(15):1-11.
[12]
ARIFI, ASLAMW, HWANGY.Barriers in adoption of internet banking:a structural equation modeling neural network approach[J].Technology in Society,2020,61:101231.
[13]
LIUM.Uav-yolo:small object detection on unmanned aerial vehicle perspective[J].Sensors,2020,20(8):2238.
[14]
XIAG S, BAIX, DINGJ, et al.DOTA:a large-scale dataset for object detection in aerial images[C]//IEEE Conferent on Computer Vision an Pattern Recognition,2018:3974-3983.
[15]
LÜX, DUANF, JIANGJ J,et al.Deep metallic surface defect detection:the new benchmark and detection network[J].Sensors,2020,20(6):1562 .