【作者】 胡骏；

【导师】 王世杰； 魏炜；

【作者基本信息】 东南大学 ， 软件工程（专业学位）， 2019， 硕士

【摘要】 近年来,计算力与数据量的增长推动了深度学习算法与模型的飞速发展,卷积神经网络作为深度学习在计算机视觉领域的基石,在图像分类、目标检测、图像语义分割等细分领域取得了一系列突破性的研究成果,其强大的特征学习能力和分类能力引起了广泛关注。一些研究者将卷积神经网络应用到犬种识别问题上,期望改善传统犬种识别方法在成本、技术难度、准确度与泛化性方面的不足,但由于训练策略、训练方法与模型选择等方面仍有缺陷,现有方法的资源消耗与识别准确度距离工程应用尚有一定距离。论文中结合了现有的犬种识别研究,针对研究中的问题进行改进,使用改进后的多网络融合模型实现面向各终端用户的犬种识别系统。主要工作如下:(1)基于卷积神经网络,设计与实现了对120种常见犬类的、准确度不低于80%的、平均响应时间小于1s的犬种识别系统,是较为成熟的卷积神经网络应用。与其他使用卷积神经网络的犬种识别系统相比,本系统通过进行模型筛选与融合提高了识别的准确度,同时通过应用迁移学习等的方法重用已有的模型和通用特征,大幅度降低了实现算法的时间成本、数据资源开销与硬件资源开销,满足了工程应用的实际条件;(2)系统的后端负责运行识别算法,需要进行密集的浮点计算,其时间开销是系统的瓶颈。通过对开销的分析,结合了GPU并行运算的特性对系统的后端应用的GPU利用率进行了优化,使得一台服务器即可实现了高质量的服务,能够在日访问量不超过20万时保证访问质量,节约了搭建服务器集群的额外设备开销、维护开销与技术难度。犬种识别系统满足了用户对于120种常见犬类的分类需求,有助于犬类的饲养者与管理者快速地、便捷地、低成本地获取犬的品种信息,同时本系统通过为宠物商城、宠物零售商引入流量与投放广告获得经济效益。更多还原

【Abstract】 The growth of computation and data in recent years has promoted the development of deep learning algorithms and models.As the cornerstone of deep learning in the field of computer vision,convolutional neural networks(CNN)has achieved a series of breakthrough research results in the fields of image classification,object detection,semantic segmentation and so on.The powerful ability of CNN for feature learning and classification attracts wide attention.Some researchers have applied CNN to dog breed identification to improve the cost,technical difficulty,accuracy and generalization of traditional methods.However,due to the shortcomings of training strategy,training method and model selection,the cost and accuracy of these researches are still far from engineering application.A fusion CNN models which have higher accuracy on dog breed identification is implemented in this thesis,and be used to implement a dog breed identification system which provide predict services to different devices(PC,mobile,etc.).The main work of this thesis is as follows:(1)A dog breed identification system with higher accuracy of no less than 80% and an average response time less than 1s for 120 kinds of common dogs is designed and implemented based on convolutional neural network,it is a mature application of convolutional neural network.Compare with ohter dog breed system which mainly based on convolutional neural network,the accuracy of this system is improved by reselect and fusion CNN models and the time cost,data resource cost and hardware resource cost of this system are greatly reduced to meets the actual conditions of engineering application by reusing the existing models and common features through transfer learning method.(2)The back-end of this system is responsible for running the dog breed algorithm.Because the algorithm needs intensive computation,the time cost is the bottleneck of the system.To improve efficiency of back-end,parallel computing feature of GPU be more competently used in back-end of this system.After optimization,the system can support 200,000 daily page view by one server,which saves the additional equipment cost,maintenance cost and technical difficulty of setting up server cluster.The dog breed identification system meets the needs of dog breed for 120 breed of common dogs,and it is helpful for owners and managers to obtain dog information quickly,conveniently and cheaply.At the same time,this system can gain economic benefits by introducing customer and advertising for pet stores and pet retailers.更多还原