基于动态随机卷积神经网络的手写数字识别方法

发布时间：2018-10-31 19:53

【摘要】：图像分类识别主要是从原始图像里划分兴趣区域并进行准确分割,并在此基础上进行分类识别任务。近年来计算机视觉与模式识别特别是卷积神经网络的发展,为分类识别提供了良好的技术支持。由于图像分类识别在视频监控、人脸识别、图像分类检索等方面有着广泛的应用前景,因此越来越受到计算机视觉领域研究者的广泛关注与研究。卷积神经网络是近年发展起来,并引起广泛重视的一种图像分类方法,传统识别方法需要训练大量网络参数,造成了训练时间的增加和网络的过拟合;输入集需要进行前期预处理,丢失了图像的原有特征。与传统方法不同,卷积神经网络不需要针对特定的任务采集图像的特征,而是模拟人类的视觉系统层次化、抽象的产生分类结果,卷积神经网络创新的采用了局部感受野,权值共享,卷积采样技术,减少了网络的训练参数数量,提高了识别速度,使得其在图像识别领域得到了广泛应用。本文从神经网络的基本概念和算法入手,深入研究神经网络理论,进而研究卷积神经网络,通过阐述常见卷积神经网络的不足,在传统卷积神经网络上修改网络结构,提出了基于动态随机卷积神经网络,并基于此理论进一步开展手写数字识别方向的研究,最后通过实验验证其网络模型的有效性和实用性。论文的主要工作如下:(1)整理和总结了近年来阐述了图像识别的研究背景和国内外研究现状,特别是卷积神经网络的国内外发展现状,介绍了神经网络和卷积神经网络的基本概念,详细阐述了网络框架和网络参数,包括网络的卷积层,池化层和梯度下降训练方法。(2)针对传统卷积神经网络对于原始图像大小的局限性,本文提出了一种动态随机卷积神经网络结构和一种随机池化方法,避免了原始图像大小的局限性,更大程度了保留了图像的纹理特征和局部特征。实验结果表明,改进的卷积神经网络精度优于传统卷积神经网络。(3)对全文做总结,提出了自身的不足和未来的研究方向。
[Abstract]:Image classification and recognition is mainly based on dividing the region of interest from the original image and accurately segmenting, and on the basis of this, the task of classification and recognition is carried out. In recent years, the development of computer vision and pattern recognition, especially convolution neural network, provides a good technical support for classification recognition. Because image classification and recognition have wide application prospects in video surveillance, face recognition, image classification and retrieval, more and more researchers in the field of computer vision pay more and more attention to it. Convolution neural network is a kind of image classification method which has been developed in recent years and has attracted wide attention. Traditional recognition methods need to train a large number of network parameters, which results in the increase of training time and the over-fitting of network. The input set needs to be preprocessed and the original feature of the image is lost. Different from the traditional methods, the convolution neural network does not need to capture the characteristics of the image for a specific task, but simulates the human visual system hierarchy, abstractly produces the classification result, the convolutional neural network innovatively adopts the local receptive field. Weight sharing and convolution sampling technology reduce the number of network training parameters and improve the recognition speed. It is widely used in the field of image recognition. In this paper, the basic concept and algorithm of neural network are introduced, the theory of neural network is deeply studied, and then the convolutional neural network is studied. By expounding the deficiency of common convolutional neural network, the network structure is modified on the traditional convolution neural network. Based on the dynamic random convolution neural network, the recognition direction of handwritten numerals is further studied based on this theory. Finally, the validity and practicability of the network model are verified by experiments. The main work of this paper is as follows: (1) the research background of image recognition in recent years and the current research situation at home and abroad, especially the development of convolution neural network, are summarized. The basic concepts of neural network and convolutional neural network are introduced. The network framework and network parameters, including the convolution layer of the network, are described in detail. (2) in view of the limitation of traditional convolution neural network to the original image size, a dynamic random convolution neural network structure and a random pool method are proposed. The limitation of the original image size is avoided, and the texture feature and local feature of the image are preserved to a greater extent. The experimental results show that the accuracy of the improved convolution neural network is better than that of the traditional convolution neural network. (3) the paper summarizes the full text and puts forward its own shortcomings and future research directions.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】