【作者】 王海龙；

【导师】 屈桢深；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2018， 硕士

【摘要】 随着视频监控应用的普及和计算机科学与技术的快速发展,计算机视觉技术得到了广泛的应用。基于视觉的目标检测技术是计算机视觉领域中非常重要的一部分,许多更高层次的视觉处理和分析任务也是以此为基础。目标检测技术被广泛应用于人脸识别、智能驾驶等领域。传统的目标检测算法基于人工提取特征,准确率容易受采集条件等因素影响,且效率低、泛化性弱。近年来基于深度学习的目标检测算法在检测准确率上远远领先传统目标检测算法,背靠大数据自动学习提取特征,效果远优于人工设计特征。虽然深度学习在许多方面都远超传统算法,但是其缺点也非常明显,需要大量的训练数据,模型参数量大,运算速度慢,对硬件要求极高。由于这些原因,深度学习算法很难做到实时,很难应用到嵌入式或小型设备中。本文研究在保证准确率的前提下,加快基于深度学习的目标检测算法的检测速度。本文具体工作内容如下:首先,详细介绍了主流CNN网络模型,分析对比了流行的基于深度学习的目标检测算法的模型结构及其它们的优劣势。然后,借鉴流行的one-stage方法,设计了直接对目标进行分类和位置回归的端到端网络模型。并使用Tensor Flow平台训练网络,使得网络的m AP达到0.713,检测速度在GTX 1080上达到32.4FPS。最后,在设计的卷积神经网络的基础上,用提取特征能力更强的Res Net网络替换VGG基网络,提升了模型检测准确率。使用K-means++方法对数据集目标框的长宽比进行聚类,然后对网络的锚框进行改进,提升了网络的检测速度。接着使用分离卷积对网络模型进行压缩,使用卷积分离的残差块重构基网络,使得网络的参数量和计算量下降到原来的1/8~1/9,网络的m AP达到0.727,检测速度达到47FPS。实验结果表明,本文对加快深度神经网络的计算速度,增强其实时性有一定的价值。更多还原

【Abstract】 With the popularization of video surveillance applications and the rapid development of computer science and technology,computer vision technology has been widely used.Vision-based object detection technology is a very important part in the field of computer vision,and many higher-level visual processing and analysis tasks are based on this.Object detection technology is widely used in face recognition,smart driving and other fields.The traditional object detection algorithm is based on artificial extraction features,the accuracy rate is easily affected by factors such as the acquisition conditions,and the efficiency is low and generalization is weak.In recent years,the object detection algorithm based on deep learning has far exceeded the traditional object detection algorithm in detection accuracy,and backed by big data to automatically learn and extract features,the effect is far better than the artificial design feature.Although deep learning is far more than traditional algorithms in many aspects,its disadvantages are also very obvious.It requires a large amount of training data,large model parameters,slow operation speed,and extremely high hardware requirements.For these reasons,deep learning algorithms are difficult to implement in real-time and are difficult to apply to embedded or small devices.This paper studies the speed of detection based on deep learning based on the premise of ensuring accuracy.The specific work of this article is as follows:First,the mainstream CNN network model is introduced in detail,and the model structure of the popular object detection algorithm based on deep learning and their advantages and disadvantages are analyzed and compared.Then,referring to the popular one-stage method,we design an end-to-end network model that directly classifies and regresses the target.And using the Tensor Flow platform to train the network,the network’s m AP reaches 0.713,and the detection speed reaches 32.4 FPS on the GTX 1080.Finally,based on the convolutional neural network designed,the VGG-based network is replaced by the Res Net network with stronger feature extraction capability,which improves the model detection accuracy.The K-means++ method is used to cluster the aspect ratio of the target frame of the data set,and then the anchor frame of the network is improved,which improves the detection speed of the network.Next,the network model is compressed using a deconvolution,and the base network is reconstructed using a residual block of convolutional separation.As a result,the network’s parameter amount and calculation amount are reduced to 1/8 to 1/9,and the network’s m AP reaches 0.727.Detection speed reaches 47 FPS.The experimental results show that this paper has certain value to speed up the calculation speed of deep neural network and enhance its real-time performance.更多还原