基于深度学习的目标识别研究

发布时间：2018-01-12 14:28

本文关键词：基于深度学习的目标识别研究　出处：《南昌航空大学》2017年硕士论文　论文类型：学位论文

【摘要】：对图像中的目标精确识别技术已应用在工业生产的各个方面,特别在军事,金融,高端装备,交通安全等领域,其关键技术在于如何提高图像识别的精准度以及识别速度,这直接影响了该技术的实用性和安全性。深度学习是机器学习领域的一个分支,本质是一种无监督学习算法,深度学习在自然语言处理,自然图像特征学习方面的效果远远超过了传统机器学习相关技术,传统的图像识别算法都是人工加入特定特征来识别目标的,不仅识别率低,而且提取困难,由于神经网络在架构上的特殊性,使得网络在训练过程中能够学习到样本中的内在规律,我们称之为样本特征,这种高度的抽象能力使得训练出来的网络具有极高的泛化能力。本文围绕着深度学习在目标识别方面的应用,主要展开了以下研究工作:(1)阐述了神经网络模型的基本结构,其中卷积层与池化层交叉堆叠可以学习抽象的特征,降采样层可以在最大程度保留原图信息的同时减少数据处理量,论文还详细阐述了反向传播算法。(2)介绍了一种特殊的无监督深度神经网络——深度自编码器。其输入输出向量具有相同维度,常按照输入向量的某种形式,通过隐层学习一个数据的表示或对原始数据进行有效编码。(3)提出了一种选择性搜索结合二值规范化梯度的目标检测策略,对比传统的滑动模型,实验证明本文提出的策略较其他算法能够提高6%的检测精度,达到相同检测率和召回率时能减少约1500个预选区域窗口。
[Abstract]:The technology of accurate target recognition in images has been applied in all aspects of industrial production, especially in military, finance, high-end equipment, traffic safety and other fields. The key technology is how to improve the accuracy and speed of image recognition, which directly affects the practicability and security of the technology. Depth learning is a branch of machine learning. Nature is an unsupervised learning algorithm, depth learning in natural language processing, natural image feature learning effect far more than the traditional machine learning related technology. Traditional image recognition algorithms are artificial to add specific features to identify the target, not only low recognition rate, but also difficult to extract, because of the special structure of neural network. So that the network in the process of training can learn the inherent laws in the sample, we call it sample characteristics. This kind of high abstract ability makes the trained network have extremely high generalization ability. This paper focuses on the application of depth learning in target recognition. The basic structure of the neural network model is expounded, in which the convolution layer and the pool layer cross stack can learn abstract features. The demultiplexing layer can keep the original image information to the maximum extent and reduce the data processing amount at the same time. This paper also introduces a special unsupervised depth neural network-depth self-encoder, whose input and output vectors have the same dimension. According to some form of input vector, a target detection strategy based on selective search combined with binary normalized gradient is proposed by learning the representation of a data or effectively encoding the original data by hidden layer. Compared with the traditional sliding model, the experimental results show that the proposed method can improve the detection accuracy of 6%, and reduce about 1 500 pre-selected region windows when the same detection rate and recall rate are achieved.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】