摘 要：本文针对视频图像中显著性检测存储量大、计算复杂等特点，提出了一种结合卷积的Hopfield 神经网络的深度学习模型。利用Hopfield 网络在序列图像处理上的优势，处理视频信息中的能更好地结合上下文关系和时序信息提取运动目标的特征。将Hopfield 的循环反馈与卷积结合起来，以在空间上更好地提取运动目标。将传统的Hopfield 网络的全连接转换为局部连接的Hopfield 网络。用局部连接的HNN 作为门控RNN 的主要部分代替区域框，并与卷积神经网络结合起来进行显著性检测，然后结合到darknet 框架下进行视频运动目标检测。在VIVD 数据集下验证显示，针对视频中的运动目标，在无须提前训练和标记的情况下能获得较好的检测结果。
关键词：局部连接；卷积Hopfield 神经网络；运动目标检测；视频图像
中图分类号：TP18；TP391.41 文献标识码：A 文章编号：2096-4706（2019）09-0074-04
Moving Target Detection Model Based on Convolutional Hopfield Network
SHEN Hui，WANG Senmei
（South-central University for Nationalities，Wuhan 430074，China）
Abstract：A deep learning model of Hopfield neural network combined with convolution was proposed for the large storage capacity and complex calculation of significance detection in video images. With the advantage of Hopfield network in processing sequential images，video information can be better combined with context and temporal information to extract features of moving targets. The circular feedback of Hopfield is combined with convolution to better extract moving targets in space. Meanwhile，the traditional Hopfield network full connection is transformed into a locally connected Hopfield network in this paper. The locally connected HNN was used as the main part of the gated RNN to rep lace the region frame and was combined with the convolutional n eural network for significance detection，and then combined with darknet for video moving target detection. In the case of VIVD data set，the verification shows that for the moving targets in video，better detection results can be obtained without the need of training and marking in advance.
Keywords：local connection；convolutional Hopfield neural network；moving object detection；video image
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[C].International Conference on Neural Information Processing Systems.2015
[2] Lin Z，Yuan C.A Very Deep Sequences Learning Approach for Human Action Recognition[C].International Conference on Multimedia Modeling.Springer，Cham， 2016.
[3] Wang P，Li W，Gao Z，et al.Action Recognition from Depth Maps Using Deep Convolutional Neural Networks[J].IEEE Transactions EgTest01 EgTest02 PkTest01 PkTest02 on Human-Machine Systems，2016，46（4）：498-509.
[4] Scesa V，Henaff P，Ouezdou FB，et al.Onthe Analysis of Sigmoid Time Parameters for Dynamic Truncated BPTT Algorithm[C].International Joint Conference on Neural Networks.IEEE，2006.
[5] John J. Hopfield. Understanding Emergent Dynamics：Using a Collective Activity Coordinate of a Neural Network to Recognize Time-Varying Patterns[J].Neural Computation，2015：1-28.
[6] Pajares G.A Hopfield Neural Network for Image Change Detection[M].IEEE Press，2006.
[7] Redmon J，Divvala S，Girshick R，et al.You Only Look Once：Unified，Real-Time Object Detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition（CVPR）. IEEE Computer Society，2016.
[8] GU F Q，HONG Z Z，SUN Y S. Boundary extraction based on stack filter，Hopfield neural network and self-organization neural network [C]// Machine Learning and Cybernetics，2003 International Conference on. S.l.：s.n.，2003：1084-1087.
[9] Narayana，Manjunath，Hanson，Allen，Learned-Miller，Erik. Coherent Motion Segmentation in Moving Camera Videos Using Optical Flow Orientations[C]. S.l.：s.n.，2013：1577-1584.
[10] Hu Y ，Chen Z，Chi Z，et al.Learning to detect saliency with deep structure[C].IEEE International Conference on Systems. IEEE，2016.
[11] Rastegari M，Ordonez V，Redmon J，et al.XNORNet：ImageNet Classification Using Binary Convolutional Neural Networks[C].European Conference on Computer Vision，2016.
作者简介：
沈慧（1994-），女，汉族，河南人，硕士研究生，研究方向：视觉计算与图像处理；
通讯作者：
王森妹（1995-），女，白族，云南人，硕士研究生，研究方向：视觉计算与图像处理。