基于结构化剪枝的遥感飞机检测算法

王成龙; 赵倩; 赵琰; 郭彤

doi:doi:10.3969/j.issn.1671-637x.2022.06.008

电光与控制, 2022, 29 (6): 37, 网络出版: 2022-08-01

基于结构化剪枝的遥感飞机检测算法

Remote Sensing Aircraft Detection Algorithm Based on Structural Pruning

论文大纲

王成龙赵倩赵琰郭彤

作者单位

上海电力大学电子与信息工程学院, 上海 201000

遥感检测飞机锚框聚类模型剪枝 remote sensing detection aircraft anchor frame clustering model pruning 

摘要

针对在遥感飞机目标检测场景中轻量级算法难以兼顾准确性与实时性的问题, 提出了一种基于YOLOv4结构化剪枝的模型压缩方法。为了使锚框参数更加符合遥感数据集并发挥网络多尺度检测的优势, 采用K-means++算法对数据集进行锚框聚类分析, 并设计尺度自适应调整, 抑制小目标过多以及目标大小接近造成的锚框冗余。此外, 为了减少模型参数量, 利用归一化层中的缩放因子γ进行L1稀疏正则化, 重新评估滤波器及卷积核权重, 对特征信息较少的通道进行迭代剪枝, 然后微调剪枝模型恢复精度。实验结果表明, 剪枝后模型参数量压缩了93.1%, 检测速度比原模型提升2.46倍, 能够在保证检测准确性的同时有效提升检测实时性。

Abstract

Regarding the problem that lightweight algorithm in remote sensing aircraft target detection is difficult to balance accuracy and real-time performancea model compression method based on YOLOv4 structured pruning is presented.In order to make the anchor frame parameters more suitable for remote sensing datasets and take advantage of network multi-scale detectionK-means++ algorithm is used to cluster the datasets and scale adaptive adjustment is designed to restrain the redundancy of the anchor frame caused by too many small targets and close target sizes.In additionin order to reduce the parameters of the modelthe scaling factor γ in the normalization layer is used for L1 sparse regularizationthe filter and convolution kernel weights are re-evaluatedchannels with less feature information are iteratively prunedand then the pruning model is fine-tuned to recover accuracy.The experimental results show that after pruningthe model parameters are compressed by 93.1%and the detection speed is 2.46 times faster than that of the original modelwhich can effectively improve the detection accuracy and real-time performance.

参考文献

[1] YA YPAN RJING Z Let al.Fusion object detection of satellite imagery with arbitrary-oriented region convolutional neural network［J］.Aerospace Systems20192:163-174.

[2] DENG Z PSUN HZHOU S Let al.Multi-scale object detection in remote sensing imagery with convolutional neural networks［J］.ISPRS Journal of Photogrammetry and Remote Sensing2018145:3-22.

[3] XU D QWU Y Q.Improved YOLO-V3 with DenseNet for multi-scale remote sensing target detection［J］.Sensors 202020(15):4276.

[4] HUANG GLIU ZMAATEN L V Det al.Densely connected convolutional networks［C］//IEEE Conference on Computer Vision and Pattern Recognition (CVPR).HonoluluHI:IEEE2016:2261-2269.

[5] REDMON JFARHADI A.YOLOv3:an incremental improvement［R］.Seattle:University of Washington2018.

[6] 公明,刘妍妍,李国宁.改进YOLO-v3的遥感图像舰船检测方法［J］.电光与控制,2020,27(5):102-107.

[7] CHENG Z BZHANG F Q.Flower end-to-end detection based on YOLOv4 using a mobile device［J］.Wireless Communications and Mobile Computing20202020:1-9.

[8] BOCHKOVSKIY AWANG C YLIAO H Y M.YOLOv4:optimal speed and accuracy of object detection［J］.Ar Xivabs/2004.10934:1-17.

[9] LIU ZLI JSHEN Zet al.Learning efficient convolutional networks through network slimming［C］//International Conference on Computer Vision (ICCV).Venice:IEEE 2017:2755-2763.

[10] 刘杰,朱旋,宋密密.改进YOLOv2的端到端自然场景中文字符检测［J］.控制与决策,2021,36(10):2483-2489.

[11] AGGARWAL SSINGH P.CuckooBat and Krill Herd based k-means++ clustering algorithms［J］.Cluster Computing201922:14169-14180.

[12] HU JSHEN LALBANIE Set al.Squeeze-and-excitation networks［J］.IEEE Transactions on Pattern Analysis and Machine Intelligence202042(8):2011-2023.

[13] HE K MZHANG X YREN S Qet al.Spatial pyramid pooling in deep convolutional networks for visual recognition［J］.IEEE Transactions on Pattern Analysis and Machine Intelligence201537(9):1904-1916.

[14] IOFFE SSZEGEDY C.Batch normalization:accelerating deep network training by reducing internal covariate shift［C］//The 32nd International Conference on International Conference on Machine Learning.Lille:IEEE 2015:448-456.

[15] 李坚强,王成,黄志超,等.基于集成知识蒸馏的肺病分析［J］.控制理论与应用2021,38(1):130-136.

[16] ZHENG Z HWANG PLIU Wet al.Distance-IoU loss:faster and better learning for bounding box regression［C］//Proceedings of the AAAI Conference on Artificial Intelligence.［S.l.］:AAAI2020:12993-13000.

[17] 姚群力,胡显,雷宏.基于多尺度卷积神经网络的遥感目标检测研究［J］.光学学报,2019,39(11):346-353.

[18] LIU WANGUELOV DERHAN Det al.SSD:single shot multi box detector［C］//European Conference on Computer Vision.Cham:Springer2016:21-37.

[19] 邵伟平,王兴,曹昭睿,等.基于MobileNet与YOLOv3的轻量化卷积神经网络设计［J］.计算机应用,2020,40(增刊1):8-13.

王成龙, 赵倩, 赵琰, 郭彤. 基于结构化剪枝的遥感飞机检测算法[J]. 电光与控制, 2022, 29(6): 37. WANG Chenglong, ZHAO Qian, ZHAO Yan, GUO Tong. Remote Sensing Aircraft Detection Algorithm Based on Structural Pruning[J]. Electronics Optics & Control, 2022, 29(6): 37.

基于结构化剪枝的遥感飞机检测算法

关于本站 Cookie 的使用提示

全站搜索

基于结构化剪枝的遥感飞机检测算法

相关论文

相关资讯

关于本站 Cookie 的使用提示

全站搜索