基于卷积神经网络的图像识别方法研究

发布时间：2019-02-24 14:44

【摘要】：随着软硬件技术的飞速发展和大数据时代的到来,深度学习技术在越来越多的领域中取得了突破性成果。而其中的卷积神经网络受生物视觉系统启发,将深度学习思想引入到神经网络之中,依靠自身具备的局部感受野、参数共享、降采样和稀疏连接等特点,近年来已成为计算机视觉领域和人工智能领域中相对活跃的研究方向。本文结合图像的自身特点,在卷积神经网络理论的基础上,论述了用于图像分类识别任务的卷积神经网络模型的构造方法、迁移训练、优化算法和分类识别性能,其主要的工作内容如下:(1)多数识别方法常将图像识别任务分割为特征提取和分类识别两个阶段,识别过程中往往涉及到繁杂的图像预处理和人工标记特征设计,考虑到需要充分保留图像的原始特征信息,利用卷积神经网络构建了一个11层端到端的图像分类识别模型。该识别模型可以从原始图像中进行特征自学习,设计的分层结构将传统方法中的特征提取和分类识别两个阶段串联融合在一起,可以进行端到端的分类识别。最后将该模型应用到自然场景下的车牌字符识别任务上,对比实验验证了该卷积模型集特征自学习和高效分类识别于一体的应用性能。(2)针对深度卷积网络训练过程中常出现的过拟合问题,构造了若干个具有不同特征提取层的深度卷积模型,分析讨论了图像数据增强变换、卷积核的尺寸、梯度下降训练优化算法以及Dropout丢弃层等模型调参方法对过拟合问题的影响,在不同参数的模型训练损失对比实验上,验证了最优识别模型的抗过拟合性。(3)针对训练数据不足的小样本量图像识别问题,提出了基于卷积神经网络的迁移学习方法。借助大规模数据集上的预训练模型,根据源数据域与目标数据域数据内容的差异,将源数据域中的预训练模型分别迁移到域内同构和域间异构的小规模图像数据集上,冻结不同层的训练参数,微调整个网络,并在部分小样本量的公共数据集和自建小规模车牌字符数据集上验证了提出方法的实验效果。
[Abstract]:With the rapid development of software and hardware technology and the arrival of big data era, the in-depth learning technology has made a breakthrough in more and more fields. The convolutional neural network is inspired by the biological vision system, and the deep learning idea is introduced into the neural network, which depends on its own characteristics of local receptive field, parameter sharing, de-sampling and sparse connection, etc. In recent years, it has become a relatively active research direction in the field of computer vision and artificial intelligence. Based on the theory of convolution neural network, this paper discusses the construction method, migration training, optimization algorithm and classification recognition performance of the convolution neural network model for image classification and recognition. The main work is as follows: (1) most recognition methods often divide the image recognition task into two stages: feature extraction and classification recognition. The recognition process often involves complicated image preprocessing and artificial label feature design. Considering the need to fully retain the original feature information of the image, an 11-layer end-to-end image classification and recognition model is constructed by using convolution neural network. The recognition model can be used for feature self-learning from the original image. The designed hierarchical structure combines the two stages of feature extraction and classification recognition in the traditional method in series and can be used for end-to-end classification recognition. Finally, the model is applied to the license plate character recognition task in the natural scene. The experimental results show that the convolutional model can be used to integrate feature self-learning and efficient classification recognition. (2) the over-fitting problem in the training process of deep convolution network is discussed. Several depth convolution models with different feature extraction layers are constructed, and the effects of image data enhancement transformation, size of convolution kernel, gradient descent training optimization algorithm and Dropout dropping layer on overfitting are discussed. In contrast experiments of model training loss with different parameters, the anti-overfitting of the optimal recognition model is verified. (3) aiming at the problem of small sample size image recognition with insufficient training data, a transfer learning method based on convolution neural network is proposed. According to the difference between source data domain and target data domain data content, the pre-training model in source data domain is migrated to small scale image data set with isomorphism and heterogeneity in domain, according to the pre-training model on large-scale data set. The training parameters of different layers are frozen, the whole network is fine-tuned, and the experimental results of the proposed method are verified on some common data sets with small sample sizes and self-built small license plate character data sets.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】