【作者】 张洁；

【导师】 赵红东；

【作者基本信息】 河北工业大学 ， 电子科学与技术， 2019， 硕士

【摘要】 随着我国汽车保有量的飞速增长与视频监控技术的广泛应用,细粒度车型识别作为智能交通系统的关键组成部分,受到国内外研究者的广泛关注。细粒度车型图像种类之间特征差异小,外观表现形式多样。同时,图像中包含的环境、光照、噪声等因素容易对特征的提取和识别造成干扰。因此,在复杂背景下精细识别汽车品牌型号面临更多难题。不同于传统方法依靠人工设计特征,卷积神经网络(CNN)包含成千上万的参数,以数据驱动的方式自主训练,具有更强的特征提取和表征能力。本文以复杂背景下的细粒度车型为分类对象,开展了基于深度卷积神经网络(DCNN)的识别研究。完成的主要工作如下:采用批量标准化、梯度下降加速算法和损失函数来优化深度卷积神经网络。深入研究卷积神经网络的结构、特点和训练过程,构建了一个十层的细粒度车型识别网络和一个包含27种相近车型的数据库。同时,从三个方面展开深度卷积神经网络的优化,并验证优化策略的可行性:通过批量标准化,使数据分布服从标准正态分布,避免内部数据的传递偏移,降低了网络训练的复杂性;采用加速优化算法,避免梯度下降陷入局部最优,使得实际输出更接近真实值;通过对比分析均方差损失、指数损失和交叉熵损失对网络训练效果的影响,选择合适的目标函数,提高网络的收敛速度。在ReLU函数的负输入部分增添含有可自适应学习参数的表达式,使激活具有负值输出,提高了深度卷积神经网络对负输入的表征能力。改进得到的EPReLU函数使各激活层的单元平均激活值更接近于零,有效减少了数据的层间传递偏移,加速了网络的收敛。识别27种细粒度车型,使用该函数的DCNN达到了98.14%的准确率。结合支持向量机(SVM)改进DCNN的Softmax层,避免了以最小化风险为目标所导致的过拟合。采用DCNN作为最初的车型识别训练模型,当训练准确率达到识别阈值后,提取其中的FC1层特征,输入SVM完成进一步训练。这样既保持了DCNN的特征提取优势,同时又避免了Softmax对已经正确分类样本的过度训练。在复杂背景下的细粒度车型识别中,通过可视化DCNN的输出特征图与卷积核权重,更加形象说明DCNN的特征学习能力。相比于SVM、Bo F和未经改进的DCNN三种识别方法,Softmax-SVM对精细车型的识别精度分别提升了54.04%、10.19%和2.27%,达到了97.78%。在用时方面,Softmax-SVM只有DCNN所耗时间的31%。更多还原

【Abstract】 With the rapid growth of automobile ownership and the wide application of video surveillance technology,fine-grained vehicle recognition,as a key component of intelligent transportation system,has attracted extensive attention from researchers.Fine-grained vehicle models have small differences in types and variety of appearances.At the same time,the environment,illumination,noise and other factors contained in the image are easy to interfere with the feature extraction and recognition.Therefore,identifying fine-grained vehicle in a complex background faces more challenges.Unlike traditional methods that rely on artificial design features,Convolutional Neural Networks(CNN),which contains thousands of parameters,is trained autonomously in a data-driven way,with enhanced feature extraction and characterization capabilities.In this paper,the recognition of fine-grained vehicle models under complex background is studied based on deep convolutional neural network.The main tasks completed are as follows:Deep convolutional neural networks is optimized by batch normalization,gradient descent acceleration algorithm and loss function.On the basis of in-depth study on the structure,characteristics and training process,a fine-grained vehicle recognition network with 10 layers and a dataset containing 27 similar vehicle types are constructed.Meanwhile,the research optimizes the deep convolutional neural networks from three aspects,and verifies the feasibility of the optimization strategy in the process of fine-grained vehicle identification.Using batch normalization,the data distribution obeys the standard normal distribution,which avoids the deviation of internal transmission and reduces the complexity of network training.Due to the adaptive optimization algorithm,the gradient descent algorithm is prevented from falling into the local optimal,so that the output is closer to the real value.Comparing and analyzing the influence of square-loss,exp-loss and cross-entropy-loss on the training effect,the appropriate objective function is selected to speed up convergence.In the negative part of ReLU,an expression containing adaptive learning parameters is added to activate the output with negative values,which enhances the expression ability of the deep convolutional neural network to the negative input.The improved function EPReLU pushes mean unit activations closer to zero,effectively reduces the bias shift of data between layers and improves the learning speed of deep neural networks.Identifying27 types of fine-grained vehicle models,the accuracy of DCNN with this function reachs98.14%.Softmax layer in DCNN is improved by combining support vector machine(SVM),that avoids over-fitting for the purpose of minimizing risk.DCNN is adopted as the initial model for vehicle models recognition and training.When the training accuracy reachs the recognition threshold,FC1 layer features are extracted and input into SVM to complete further training.The classifier can not only maintain the feature extraction advantage of DCNN,but also avoid the overtraining of Softmax on correctly classified samples.In fine-grained vehicle recognition under complex background,the ability of feature learning of DCNN is more vividly illustrated by visualizing the output feature maps and convolution kernels weight.It is found that compared with SVM,Bo F and unimproved DCNN,the fine vehicle types identification precision of softmax-svm reachs 97.78%,improving by 54.04%,10.19% and 2.27%,respectively.In terms of time,softmax-svm takes only 31% of DCNN.更多还原