基于卷积神经网络的车辆检索方法研究

The Research of Vehicle Image Retrieval Based on Convolutional Neural Network

为了解决传统车辆检索方法中准确性和区分度较低的问题,本文提出了一种基于卷积神经网络(CNN)的车辆检索方法。该方法利用CNN稀疏连接和权值共享的优点,针对车辆前脸图像关键特征位置的相对不变性,设计了一个七层的网络结构,可以合理提取车辆的有效特征,并将低级结构特征组合成为高一级的特征,既简化了模型的复杂度,也克服了旋转平移等因素对检测结果的影响。该方法最终通过相似度排序的方法得到检索结果。实验结果表明,本文所提出的方法相对于基于局部不变特征方法具有更高的准确度。

In order to solve the problem of the traditional vehicle retrieval methods which has low degree of differentiation and accuracy, it is proposed in this paper that a vehicle retrieval method based on convolution neural network(CNN).This methods builds a seven layers network structure which utilizes the advantages of CNN which called sparse connecting and weights sharing, and the key feature location invariance for vehicle former face image. This structure can extract effective features of vehicle which is reasonable, and combine the low scale feature into higher scale one. It not only simplifies the complexity of the model,but also overcome the influence of rotating shift on the result with effect. Finally, it get the retrieve result through similarity ranking. By comparing with the methods which based on local invariant features, the experimental results show that the presented method has a higher accuracy.