基于预决策的卷积神经网络加速研究

发布时间：2018-01-13 16:01

本文关键词：基于预决策的卷积神经网络加速研究　出处：《北京工业大学》2016年硕士论文　论文类型：学位论文

【摘要】：卷积神经网络(Convolutional Neural Networks,CNN)目前在计算机视觉领域应用广泛,在图像识别、物体检测以及图像分割等问题上都取得了最优的性能。但是CNN是深度的人工神经网络,其前馈传播需要进行大量的卷积运算,而且其网络层数和隐藏层的网络节点数会随着任务难度的增加而增加。因此,CNN过高的计算复杂度,对其在实时性要求较高的应用场景中的使用产生了限制,如对行人和车辆进行检测的交通监控等问题;对其在计算能力较低的设备上的应用也有很大的限制,如移动设备等。本文主要对CNN在图像识别和物体检测中应用的加速方法展开了系统而深入的研究,主要包括以下内容:1.针对CNN在二元图像分类和特定物体检测问题上的应用,提出了一种基于预决策的卷积神经网络加速方法。该方法充分利用了CNN特征的冗余性和多层特征的判别力,建立了基于CNN的级联分类器,实现加速。在构建级联分类器的过程中,对于每一阶的特征选择,提出了特征点选择模型。特征点选择模型充分考虑了CNN各层特征点的分类能力和计算代价,跨层进行特征点的选择,所选特征点的分类能力和计算代价整体最优。本文分别在CIFAR-10、Pascal VOC2007、INRIA和TRECVID等数据集上对该方法进行了测试。实验表明,对于二元图像分类和特定物体检测问题,该方法在有极小性能损失的情况下,能实现较大的加速。另外,该方法可以与很多之前的CNN加速方法结合起来使用,实现更大的加速。2.对基于预决策的卷积神经网络加速方法进行了系统实现。该系统主要包括:特征点选择模型训练集提取模块、CNN网络结构量化模块、特征点选择模型训练模块、级联分类器建立模块。对于深度较大和比较复杂的CNN模型,本文提出的方法在训练过程中时间和空间复杂度都非常大,该系统使用了一些实现技巧,大大压缩了模型的训练时间和所需内存空间。
[Abstract]:Convolutional Neural Networks (CNN) is widely used in the field of computer vision, and is widely used in image recognition. Object detection and image segmentation have achieved optimal performance, but CNN is a depth of artificial neural network, its feedforward propagation requires a large number of convolution operations. Moreover, the number of network layers and the number of network nodes in the hidden layer will increase with the increase of task difficulty. It has some limitations on the use of real-time applications, such as traffic monitoring for pedestrian and vehicle detection. There are also some limitations to the application of CNN in devices with low computing power, such as mobile devices. In this paper, the acceleration method of CNN in image recognition and object detection is studied systematically and deeply. The main contents are as follows: 1. Aiming at the application of CNN in binary image classification and specific object detection. This paper presents an accelerated method of convolution neural network based on pre-decision, which makes full use of the redundancy of CNN features and the discriminant power of multi-layer features, and establishes a cascade classifier based on CNN. Implementation of acceleration. In the process of constructing cascaded classifiers, the feature selection for each order. A feature point selection model is proposed. The feature point selection model takes into account the classification ability and computational cost of each layer of CNN feature points and selects feature points across layers. The classification ability and computational cost of the selected feature points are globally optimal. The method is tested on the data sets of INRIA and TRECVID. Experiments show that the method has minimal performance loss for binary image classification and specific object detection problems. In addition, this method can be used in conjunction with many previous CNN acceleration methods. To achieve greater acceleration. 2. A systematic implementation of the convolution neural network acceleration method based on pre-decision is carried out. The system mainly includes: feature selection model training set extraction module and CNN network structure quantization module. Feature selection model training module, cascade classifier building module. For the deeper and more complex CNN model, the proposed method in the training process in the time and space complexity is very large. The system uses some implementation techniques to greatly reduce the training time and memory space required by the model.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183

【相似文献】