【作者】 田超；

【导师】 黄朝兵；

【作者基本信息】 武汉理工大学 ， 电子科学与技术， 2019， 硕士

【摘要】 场景识别是机器视觉领域中一个非常重要的课题,随着深度学习的发展,场景识别也逐步联合深度学习应用于无人机、无人驾驶、安防监控等领域中,且取得了良好的识别效果,基于深度学习的场景识别也成为场景识别领域中的重点研究方向。场景识别可以分为室内场景识别和室外场景识别,室外场景一般由开阔的背景信息组成,识别难度比室内场景识别难度高。对于无人机、无人驾驶汽车以及室外机器人来说,感知和识别周围环境十分重要,因此,研究基于深度学习的室外场景识别问题具有重要意义。本文的主要研究工作如下:(1)室外场景多由大量背景信息组成,识别难度较高,为满足本文场景识别研究要求,本文采用在ImageNet数据集上已预训练的深度网络模型。针对自建数据集和ImageNet数据集类别差异问题,采用在预训练模型的基础上进行微调的GoogleNet Inception V3网络模型,并对网络模型进行优化时的学习率的设置问题进行探讨研究。(2)深度学习反向传播中的学习率设置的过小会导致网络模型收敛速度过慢,延长网络的训练时间,过大的学习率又会导致发散,针对此问题,本文提出一种与损失函数相关的基于计算动词理论动态调整学习率的方法,优化网络模型。为描述损失函数和学习率之间的关系,本文引入最陡下降算法的最小均方(Least mean square,LMS)自适应滤波算法中步长因子和误差的关系公式,将关系公式作为学习率的衰减公式训练网络模型。再将计算动词加入到学习率与损失函数关系公式中,使用计算动词相似度推导出用于每轮优化的学习率,进行参数调整,优化网络模型,进而得到一个更小的损失函数,提高识别效果。(3)针对深度学习网络模型对低层特征局部细节描述相对较弱的问题,本文提出一种基于dColorSIFT的传统特征和深度特征融合的场景识别方法。采用dColorSIFT算法提取低层特征,将提取的特征依次经过词袋模型和空间金字塔模型的编码和池化后,与使用计算动词改进的GoogLeNet InceptionV3网络模型提取的特征进行融合,再通过SVM分类器进行分类识别。利用该方法弥补深度学习网络模型提取的低层特征局部细节描述不足的缺陷,提高场景识别准确率。更多还原

【Abstract】 Scene recognition is a very important topic in the field of machine vision.With the development of deep learning,scene recognition is gradually combined with deep learning in the fields of drones,driverless,security monitoring,etc.,and has achieved good recognition results.Scene recognition based on deep learning has also become a key research direction in the field of scene recognition.Scene recognition can be divided into indoor scene recognition and outdoor scene recognition.Outdoor scenes are generally composed of open background information,and the recognition difficulty is more difficult than indoor scene recognition.For drones,driverless cars and outdoor robots,it is important to sense and identify the surrounding environment.Therefore,it is important to study the outdoor scene recognition problem based on deep learning.The main research work of this paper is as follows:(1)The outdoor scene is composed of a large amount of background information,and the recognition difficulty is high.In order to meet the research requirements of scene recognition,this paper adopts the deep network model pre-trained on the ImageNet dataset.For the self-built dataset and ImageNet dataset category differences,the GoogleNet Inception V3 network model based on the pre-training model is used,and the learning rate setting problem when the network model is optimized is discussed.(2)The learning rate setting in deep learning backpropagation will cause the network model to converge too slowly,prolong the training time of the network,and the excessive learning rate will lead to divergence.To solve this problem,this paper proposes a The loss function is related to the method of dynamically adjusting the learning rate based on the computational verb theory to optimize the network model.In order to describe the relationship between the loss function and the learning rate,this paper introduces the relationship formula of the step factor and error in the Least Mean Square(LMS)adaptive filtering algorithm of the steepest descent algorithm,and uses the relational formula as the learning rate.The attenuation formula trains the network model.Then the computational verb is added to the relationship between learning rate and loss function.The learning verb similarity is used to derive the learning rate for each round of optimization,parameter adjustment,optimization of the network model,and then a smaller loss function to improve recognition.effect.(3)For the deep learning network model,the description of the local details of the low-level features is relatively weak.This paper proposes a scene recognition method based on the fusion of traditional features and depth features of dColorSIFT.The dColorSIFT algorithm is used to extract the low-level features,and the extracted features are encoded and pooled by the word bag model and the spatial pyramid model in turn,and then merged with the features extracted by the GoogLeNet InceptionV3 network model improved by the computational verbs,and then classified by the SVM classifier.Identification.This method is used to compensate for the shortcomings of the low-level feature local detail description extracted by the deep learning network model,and improve the scene recognition accuracy.更多还原