【作者】 张杰；

【导师】 岳有军；

【作者基本信息】 天津理工大学 ， 控制科学与工程， 2022， 硕士

【摘要】 室内场景识别是移动服务机器人为人们提供优质服务的基础,也是当今机器视觉的研究热点之一。室内场景识别主要包括布局估计、场景分类和目标检测等。由于布局估计常常受到室内复杂环境、光照阴影和杂乱物品的影响,因此,在室内场景识别任务中难度最大。同时,由于室内场景的复杂度高,室内场景分类相比于室外场景分类也更具挑战性。据此,本文开展移动服务机器人室内场景识别关键技术研究,主要研究内容如下:1.为了简化网络结构,提高布局估计效率,本文提出一种基于改进轻量网络的实时布局估计方法。该方法利用轻量级的编码解码网络,端对端直接获得分割型布局估计;针对以往联合学习方法特征利用率不高的问题,使用简化的联合学习模块,精简联合学习网络并提高特征利用率;针对数据集正负标签不平衡和布局类型分布不平衡问题,使用分割型语义迁移,提高网络训练的稳定性。在LSUN和Hedau数据集上,使用像素误差指标对结果进行评价。实验表明,本文提出的分割型布局估计方法能够快速准确获得布局估计。2.为了提高布局估计准确率,本文提出一种分步训练的边缘型布局估计方法。首先使用边缘型布局估计网络,获得粗糙的边缘型布局。针对以往联合学习方法网络结构复杂的问题,使用分步联合学习方法简化网络结构;针对边缘型布局的训练集存在的正负标签不平衡问题,使用分步语义迁移的方法提高训练稳定性。然后采用简化的粗糙边缘型布局优化方法,提高布局估计精度和优化效率。在LSUN和Hedau数据集上,使用像素误差和关键点误差指标对结果进行评价。实验表明,本文提出的分步训练的布局估计方法,能够获得更加准确的布局估计。3.为了提高室内场景分类的准确率,本文使用卷积神经网络实现端对端的室内场景分类。为缓解数据集过小带来的问题,本文采用迁移学习的训练方法,使用在大数据集上预训练得到的网络状态,初始化本文网络的部分权重,防止样本过少产生过拟合。在LSUN数据集上,使用分类准确率指标进行评价。结果表明,本文的方法分类效果更好。4.基于本文的研究成果,使用Py Qt5和Qt Designer设计了室内场景识别软件,实现分割型布局估计、边缘型布局估计和室内场景分类功能。更多还原

【Abstract】 Indoor scene recognition is the basis of mobile service robot to provide high quality service for people,and it is also one of the research hotspots of machine vision.Indoor scene recognition mainly includes layout estimation,scene classification and target detection.Because layout estimation is often affected by indoor complex environment,light and shadow,and clutter,it is the most difficult task in indoor scene recognition.At the same time,indoor scene classification is more challenging than outdoor scene classification due to the high complexity of indoor scene.Accordingly,this paper carries out research on the key technologies of indoor scene recognition of mobile service robot.The main research contents are as follows:1.In order to simplify the network structure and improve the efficiency of layout estimation,this paper proposes a real-time layout estimation method based on improved lightweight network.This method uses a lightweight encoding and decoding network to obtain the segmentation layout estimation directly from end-to-end.Aiming at the problem of low feature utilization of previous joint learning methods,a simplified joint learning module is used to simplify the joint learning network and improve feature utilization.In order to improve the stability of network training,segmentation semantic transfer is used to solve the imbalance of positive and negative labels and layout types of data sets.On LSUN and Hedau datasets,the results were evaluated using pixel error indicators.Experiments show that the proposed segmentation layout estimation method can obtain layout estimation quickly and accurately.2.In order to improve the accuracy of layout estimation,this paper proposes a stepwise training edge layout estimation method.The rough edge layout is obtained by edge layout estimation network.Aiming at the problem of complex network structure of previous joint learning methods,a stepwise joint learning method is used to simplify the network structure.Aiming at the imbalance of positive and negative labels in the edge layout training set,a stepwise semantic transfer method is used to improve the stability of training.Then a simplified rough edge layout optimization method is used to improve the accuracy and efficiency of layout estimation.On LSUN and Hedau datasets,the results were evaluated using pixel error and corner error indicators.Experiments show that the proposed method can obtain more accurate layout estimation.3.In order to improve the accuracy of indoor scene classification,this paper uses convolutional neural network to achieve end-to-end indoor scene classification.In order to alleviate the problems caused by too small data set,this paper adopts the training method of transfer learning to initialize part of the weight of the network in this paper by using the network state pre-trained on the large data set to prevent over-fitting due to too few samples.On LSUN data set,classification accuracy index was used to evaluate.The results show that the classification effect of the proposed method is better.4.Based on the research results of this paper,Py Qt5 and Qt Designer are used to design an indoor scene recognition software to achieve segmentation layout estimation,edge layout estimation and indoor scene classification.更多还原