基于分组异构卷积的轻量级目标检测网络

doi:10.11896/jsjkx.190600067

摘要/Abstract

摘要： 目前的目标检测模型存在参数量多、模型体积大及检测速度慢的缺点,不能在实时场景下应用。例如,对于自动驾驶技术,不仅需要精准的检测来保障安全,还需要实现快速检测以保证车辆的实时决策。针对以上问题,提出了一种端对端的轻量级目标检测网络FGHDet。首先,针对异构卷积HetConv逐通道卷积效率低的问题,对特征图进行分组,提出了分组异构卷积GHConv(Grouping Heterogeneous Convolution);其次,将GHConv和Fire Module组合,构建了基础模块FGH Module;最后,以FGH Mdolue为基础,搭建了端对端的轻量级目标检测网络FGHDet。FGHDet主要通过两种方法来减少参数量:1)使用1×1的卷积对特征图进行降维,减少3×3滤波器的输入通道数量;2)使用GHConv替换传统的卷积核。以KITTI数据集为实验数据,在深度学习框架Keras上完成了模型的训练和评估。实验结果表明,FGHDet在KITTI数据集上的mAP可以达到74.4%,高于Faster R-CNN的70.8%;模型检测速度为28.7FPS,优于对比模型中最快的SqueezeDet;而且该模型的大小仅为2.6MB,是Faster R-CNN模型体积的1/200。

关键词: FGHDet, KITTI, 分组异构卷积, 目标检测, 轻量级

Abstract: The current object detection model has disadvantages like large number of parameters,large size and slow detection speed,and it cannot be applied in real-time scenarios.For example,automatic driving technology requires not only accurate detection to ensure safety,but also rapid detection to ensure real-time decision-making of vehicles.For the above questions,this paper presented an end-to-end light-weight object detection network FGHDet.Firstly,grouping heterogeneous convolution(GHConv) is proposed to solve the problem of low efficiency of HetConv channel by channel convolution.Secondly,the basic module FGH Module is built by combining GHConv and Fire Module.Finally,the end-to-end light-weight object detection network FGHDet is built based on FGH Module.FGHDet reduces the amount of parameters mainly in two ways.One is to reduce the number of input channels of the 3×3 filter,and other is to replace the traditional convolution kernel with GHConv.This paper took KITTI data set as experimental data to complete the training and evaluation of the model on the deep learning framework Keras.The experimental results indicate that the mAP of FGHDet in KITTI data set can reach 74.4%,higher than 70.8% of Faster R-CNN,and the model detection speed is 28.7FPS,better than SqueezeDet,the fastest model in the comparison models.Moreover,the size of the proposed model is only 2.6MB,1/200 times the volume of the Faster R-CNN model.

Key words: FGHDet, Grouping heterogeneous convolution, KITTI, Light-weight, Object detection

中图分类号:

TP391

晏晓天, 黄山. 基于分组异构卷积的轻量级目标检测网络[J]. 计算机科学, 2020, 47(4): 108-111. https://doi.org/10.11896/jsjkx.190600067

YAN Xiao-tian, HUANG Shan. Light-weight Object Detection Network Based on Grouping Heterogeneous Convolution[J]. Computer Science, 2020, 47(4): 108-111. https://doi.org/10.11896/jsjkx.190600067

参考文献

[1]REN S,HE K,GIRSHICK R,et al.Faster r-cnn:Towards real-time object detection with region proposal networks [M]//Advances in Neural Information Processing Systems.Berlin:Springer,2015:91-98.
[2]CHENG Y,WANG D,ZHOU P,et al.A survey of model compression and acceleration for deep neural networks [J].arXiv:1710.09282,2017,[3]CHENG J,WANG P-S,LI G,et al.Recent advances in efficient computation of deep convolutional neural networks [J].Frontiers of Information Technology & Electronic Engineering,2018,19(1):64-77.
[4]JI R R,LIN S H,CHAO F,et al.A review of deep neural network compression and acceleration[J].Journal of Computer Research and Development,2018,55(9):1871-1888.
[5]MOLCHANOV P,TYREE S,KARRAS T,et al.Pruning convolutional neural networks for resource efficient transfer learning [J].arXiv:1611.06440,2016.
[6]LI H,KADAV A,DURDANOVIC I,et al.Pruning filters for efficient convnets [J].arXiv:1608.08710,2016.
[7]YU X,LIU T,WANG X,et al.On compressing deep models by low rank and sparse decomposition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2017.
[8]ZHANG D,YANG J,YE D,et al.Lq-nets:Learned quantization for highly accurate and compact deep neural networks[C]//Proceedings of the European Conference on Computer Vision (ECCV).2018.
[9]HINTON G,VINYALS O,DEAN J.Distilling the knowledge in a neural network [J].arXiv:1503.02531,2015.
[10]MISHRA A,MARR D.Apprentice:Using knowledge distillation techniques to improve low-precision network accuracy [J].arXiv:1711.05852,2017.
[11]HOWARD A G,ZHU M,CHEN B,et al.Mobilenets:Efficient convolutional neural networks for mobile vision applications [J].arXiv:1704.04861,2017.
[12]SANDLER M,HOWARD A,ZHU M,et al.Mobilenetv2:Inverted residuals and linear bottlenecks [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2018.
[13]HOWARD A,SANDLER M,CHU G,et al.Searching for mobilenetv3 [J].arXiv:1905.02244,2019.
[14]ZHANG X,ZHOU X,LIN M,et al.Shufflenet:An extremely efficient convolutional neural network for mobile devices[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2018.
[15]MA N,ZHANG X,ZHENG H-T,et al.Shufflenet v2:Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European Conference on Computer Vision (ECCV).2018.
[16]IANDOLA F N,HAN S,MOSKEWICZ M W,et al.SqueezeNet:AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size [J].arXiv:1602.07360,2016.
[17]SINGH P,VERMA V K,RAI P,et al.HetConv:Heterogeneous Kernel-Based Convolutions for Deep CNNs [J].arXiv:1903.04120,2019.
[18]LIN M,CHEN Q,YAN S.Network in network [J].arXiv:1312.4400,2013.
[19]GEIGER A,LENZ P,STILLER C,et al.Vision meets robotics:The KITTI dataset [J].The International Journal of Robotics Research,2013,32(11):1231-1237.
[20]WU B,IANDOLA F,JIN P H,et al.Squeezedet:Unified,small,low power fully convolutional neural networks for real-time object detection for autonomous driving[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops.2017.

相关文章 15

[1]	刘冬梅, 徐洋, 吴泽彬, 刘倩, 宋斌, 韦志辉. 基于边框距离度量的增量目标检测方法 Incremental Object Detection Method Based on Border Distance Measurement 计算机科学, 2022, 49(8): 136-142. https://doi.org/10.11896/jsjkx.220100132
[2]	王灿, 刘永坚, 解庆, 马艳春. 基于软标签和样本权重优化的Anchor Free目标检测算法 Anchor Free Object Detection Algorithm Based on Soft Label and Sample Weight Optimization 计算机科学, 2022, 49(8): 157-164. https://doi.org/10.11896/jsjkx.210600240
[3]	祝文韬, 兰先超, 罗唤霖, 岳彬, 汪洋. 改进Faster R-CNN的光学遥感飞机目标检测 Remote Sensing Aircraft Target Detection Based on Improved Faster R-CNN 计算机科学, 2022, 49(6A): 378-383. https://doi.org/10.11896/jsjkx.210300121
[4]	沈超, 何希平. 基于纹理特征增强和轻量级网络的人脸防伪算法 Face Anti-spoofing Algorithm Based on Texture Feature Enhancement and Light Neural Network 计算机科学, 2022, 49(6A): 390-396. https://doi.org/10.11896/jsjkx.210600217
[5]	马宾, 付永康, 王春鹏, 李健, 王玉立. 基于GDIoU损失函数的YOLOv4绝缘子高效定位算法 High Performance Insulators Location Scheme Based on YOLOv4 with GDIoU Loss Function 计算机科学, 2022, 49(6A): 412-417. https://doi.org/10.11896/jsjkx.210600089
[6]	陈永平, 朱建清, 谢懿, 吴含笑, 曾焕强. 基于外接圆半径差损失的实时安全帽检测算法 Real-time Helmet Detection Algorithm Based on Circumcircle Radius Difference Loss 计算机科学, 2022, 49(6A): 424-428. https://doi.org/10.11896/jsjkx.220100252
[7]	陈佳舟, 赵熠波, 徐阳辉, 马骥, 金灵枫, 秦绪佳. 三维城市场景中的小物体检测 Small Object Detection in 3D Urban Scenes 计算机科学, 2022, 49(6): 238-244. https://doi.org/10.11896/jsjkx.210400174
[8]	胡伏原, 万新军, 沈鸣飞, 徐江浪, 姚睿, 陶重犇. 深度卷积神经网络图像实例分割方法研究进展 Survey Progress on Image Instance Segmentation Methods of Deep Convolutional Neural Network 计算机科学, 2022, 49(5): 10-24. https://doi.org/10.11896/jsjkx.210200038
[9]	徐涛, 陈奕仁, 吕宗磊. 基于改进YOLOv3的机坪工作人员反光背心检测研究 Study on Reflective Vest Detection for Apron Workers Based on Improved YOLOv3 Algorithm 计算机科学, 2022, 49(4): 239-246. https://doi.org/10.11896/jsjkx.210200119
[10]	张侣, 周博文, 吴亮红. 基于改进卷积注意力模块与残差结构的SSD网络 SSD Network Based on Improved Convolutional Attention Module and Residual Structure 计算机科学, 2022, 49(3): 211-217. https://doi.org/10.11896/jsjkx.201200019
[11]	赫晓慧, 邱芳冰, 程淅杰, 田智慧, 周广胜. 基于边缘特征融合的高分影像建筑物目标检测 High-resolution Image Building Target Detection Based on Edge Feature Fusion 计算机科学, 2021, 48(9): 140-145. https://doi.org/10.11896/jsjkx.200800002
[12]	袁磊, 刘紫燕, 朱明成, 马珊珊, 陈霖周廷. 融合改进密集连接和分布排序损失的遥感图像检测 Improved YOLOv3 Remote Sensing Target Detection Based on Improved Dense Connection and Distributional Ranking Loss 计算机科学, 2021, 48(9): 168-173. https://doi.org/10.11896/jsjkx.200800001
[13]	龚浩田, 张萌. 基于关键点检测的无锚框轻量级目标检测算法 Lightweight Anchor-free Object Detection Algorithm Based on Keypoint Detection 计算机科学, 2021, 48(8): 106-110. https://doi.org/10.11896/jsjkx.200700161
[14]	李琳, 刘学亮, 赵烨, 纪平. 结合乐高滤波器和SSD的低光照图像融合检测方法 Low Light Image Fusion Detection Method Based on Lego Filter and SSD 计算机科学, 2021, 48(7): 213-218. https://doi.org/10.11896/jsjkx.200800127
[15]	辛元雪, 史朋飞, 薛瑞阳. 基于区域提取与改进 LBP 特征的运动目标检测 Moving Object Detection Based on Region Extraction and Improved LBP Features 计算机科学, 2021, 48(7): 233-237. https://doi.org/10.11896/jsjkx.200600131

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed