【作者】 张哲；

【导师】 于哲舟；

【作者基本信息】 吉林大学 ， 计算机技术（专业学位）， 2018， 硕士

【摘要】 目标检测是计算机视觉的一个重要分支,其目的是从静态图片中定位并识别出图中感兴趣区中的特定目标。传统的方法先对图像中的目标进行定位,然后提取出目标中所包含的特征,最后将特征传入分类器中,完成对目标的分类。其中对目标的特征提取和分类是分开进行的,在复杂的环境中这样需要有经验的研究人员对目标进行手工提取特征,使得对目标特征提取过程中工作量巨大,并且没有得到很好的效果。随着科技的进步,硬件设备性能的提高,大数据时代的到来,基于深度学习卷积神经网络对目标进行检测的方法又被重新提出来研究,它通过对输入图像进行卷积运算,直接从图像中提取到具有表征信息的特征,并通过前向传播与反向传播调整卷积神经网络运算中的参数,并且随着卷积神经网络的深入研究,使得卷积神经网络在对目标进行特征提取时,也不要求原始图片具有固定的尺寸,从而使原始图像不需要进行剪裁或压缩的操作,使其提取到的特征信息更加完整和准确。相比传统目标检测的方法,基于深度学习卷积神经网络结构可以将目标特征的提取和分类可以加载到一个神经网络结构中,进行端到端处理,使得提取到的信息更加具有表征性,同时卷积神经网络具有权值共享和稀疏连接的特点,使得训练神经网络时需要训练的参数数量大量减少,降低了神经网络中的计算量,增加了目标检测的效率。本文主要研究内容如下:(1)本文首先介绍了课题的研究背景和意义,分析了国内外对于深度学习卷积神经网络目标检测算法的研究现状。同时介绍了传统目标检测流程,以及对图像中的目标所进行的一系列操作。(2)深入理解神经网络结构,讲述了神经网络模型训练中的前向传播和反向传播算法,理解应用深度神经网络算法中所遇到的过拟合现象以及其处理方法,同时介绍了深度学习中的几种常用的优化算法,最后讲述了有关卷积神经网络的内容和其主要特点。(3)介绍本文目标检测中基于深度学习卷积神经网络的结构,并讲述了其中的对目标特征的提取,区域建议框的生成,ROI池化层以及多任务训练,最后通过改进其中对目标特征提取的网络结构来获得更具有表征信息的特征,从而提高目标检测中的准确率。在复杂背景下,当数据量越多神经网络结构越深,则越能提取到更准确的特征信息,从而提升目标检测的准确率。但是深层次的网络结构在训练中面临梯度消失和梯度扩散的问题,本文介绍了可以解决上述问题的残差神经网络结构,最终可以通过深层网络结构获得更具有表征信息的目标特征,提高了目标检测的准确率。更多还原

【Abstract】 Object detection is an important research topic in the field of computer vision.The main purpose of object detection is to locate and identify objects of interest from a static picture.In the traditional method,the object is located first,then the feature is extracted,and then the classifier is introduced into the classifier.The feature extraction and classification of targets are carried out separately.In complex environment,it is not only necessary for experienced researchers to manually extract features,but also time-consuming.It makes a huge amount of work in the process of target extraction and does not get good results.With the rapidly increasing of technology,the improvement of machine performance and the arrival of the era of big data,target detection based on the structure of convolution neural network is reintroduced.It uses the convolution operation of input image to extract features with characteristic information directly,and adjusts parameters in convolution neural network through forward propagation and back propagation.Further more,convolution neural network does not require the original image to have a fixed size when extracting feature,so that the extracted feature information is more complete and accurate.The object feature extraction and classification based on deep learning convolution neural network object detection is much better than traditional methods that it can be loaded into a network and end to end processing,and finally get the more representative features.At the same time,convolution has the characteristics of weight sharing and sparse connection.It greatly reduces both of parameter numbers and the complexity of computation,and make the algorithm more efficient.In simple terms,the main is shown below:(1)This paper first describes the background of this research,analyzes the current research status of target learning algorithm of deep learning convolution neural network at home and abroad,introduces some of the important research of traditional algorithm,and some operations on picture processing.(2)The structure of neural network and its basic knowledge are deeply understood,and forward propagation and reverse propagation in neural network are described.It also explains the handling of over-fitting phenomenon in deep neural network,and introduces several commonly used optimization algorithms in deep learning.Finally,the basic content and main characteristics of convolution neural network are introduced in detail.(3)This paper introduces the structure of target detection based on deep learning convolution neural network in the text,and describes some methods of how to extract the target feature,the generation of region proposal box,ROI pool layer and multi-task loss.Finally,by improving the network structure extracted from the features,the characteristics of the information are obtained to improve the accuracy of the object detection.In a complex background,when the data is sufficient,with the structure of the network deeper and deeper,the more accurate features can be extracted to improve the accuracy of the object detection,however the training complexity is increasing by the number of layers,because there will be a gradient elimination and a gradient explosion.The residual neural network we use can suppress these problems well,so that we can get more characteristic information characteristics and enhance the accuracy of target detection.更多还原