一种新的基于无损失函数的深度卷积神经网络的图像特征提取方法

发布时间：2018-03-10 07:24

本文选题：图像识别　切入点：图像特征提取　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：作为人工智能的重要技术之一,图像识别技术已经在社会的各个领域被广泛应用。例如:集装箱号码图像自动识别,人脸图像识别,病理图像自动识别等等。对于一个成熟的图像识别系统,其主要步骤可以分为四步:图像采集与预处理,目标图像定位和分割,图像特征提取,特征识别与分类。其中,图像特征提取是其最为重要的组成步骤。提取到的图像特征的质量直接影响着最终的分类结果。近年来,随着深度学习的发展,越来越多的图像特征提取采用了深度学习技术。由于深度卷积神经网络本身独特的结构优势,使其成为众多给予深度学习的图像识别程序中最为主流的图像特征提取手段。然而,由于目前深度学习研究趋势向着多层次,更复杂的方向,训练一个收敛的深度网络难度越来越高,不仅需要调整越来越多的参数,还需要保证其损失函数的收敛性,有时候还需要整合一些现有的训练技巧,例如dropout,maxout等。近年来,随着简单深度学习结构的提出,越来越多的研究人员着眼于这种无损失函数的深度学习模型。其中,最为著名的是PCANet深度卷积网络模型。其无损失函数的特性使得网络训练的难度大大降低。然而,由于其选择使用的无监督的主成分分析(Principal Component Analysis,PCA)算法进行卷积核的训练,使得最终的图像识别效果并不是那么理想。虽然随后基于监督学习的线性判别式分析(Linear Discriminant Analysis,LDA)方法来训练卷积核的深度学习模型LDANet被提出,但是由于LDA算法自身正负样本分离能力的局限性,使得LDANet模型的图像识别结果并无显著提升。并且,因为其特有的下采样方式,使其极易产生过度拟合现象。使得训练效果不尽如人意。在此论文中,针对PCANet与LDANet所出现的问题,主要做出了以下的贡献:1)本文基于PCANet的基本结构,创造性地将Marginal Fisher Analysis(MFA)引入卷积核的训练,提出了一种新的简单深度学习架构MFANet。由于MFA采取监督学习的方式,并且通过提取后的特征向量进行映射,使得投影后的正样本之间的距离尽可能小,同时保证负样本之间的距离尽可能大。使得卷积同类图像所得的特征与卷积后的异类图像特征在新的特征空间得以有效分离。本文运用标准数据集:字符识别数据集ICDAR2003,测试所提出的深度模型的识别性能,并且通过和一些流行的图像特征提取模型进行比较,最终的图像分类的结果表明MFANet模型的特征提取能力比其他深度模型要好。2)本文提出了一种新的下采样方法:基于概率的块级随机直方图化,来解决基于PCANet的模型所具有的鲁棒性不强的特点。该方法的核心是计算块级像素出现概率,并基于该概率去采样决定块级特征像素点是否保留,最终运用直方图输出特征。通过运用图像识别数据集:PIE FACE数据集,将所提出的下采样方法所池化的特征分类结果的与其它流行的下采样后的特征分类结果进行比较,实验结果表明本文所提出的方法有效地减少了过拟合现象。
[Abstract]:As one of the important technology of artificial intelligence, image recognition technology has been widely used in various fields of society. For example: automatic identification of container number image, face image recognition, image recognition and so on. The pathological image recognition system for a mature, the main steps can be divided into four steps: image acquisition and preprocessing, target image location and segmentation, feature extraction, feature recognition and classification. The image feature extraction is the most important component steps. The quality of image feature extraction to directly affect the final classification result. In recent years, with the development of deep learning, image feature extraction has been used increasingly deep learning technology due to structural advantage of deep convolutional neural network is unique, which give as many image features most mainstream image recognition program to extract deep learning means. However, due to the current trend towards multi-level deep learning research, more complex direction, depth of network is more and more high training a convergence, not only need to adjust the parameters of more and more, but also need to ensure the convergence of the loss function, sometimes also need to integrate some of the existing training techniques, such as dropout, maxout etc. in recent years. With the development of deep learning, simple structure, focus on the deep learning model to the loss function more and more researchers. Among them, the most famous is the PCANet network model. The depth of the convolution loss function makes the network training difficulty is greatly reduced. However, due to the choice of unsupervised use of principal components the analysis (Principal Component Analysis PCA) algorithm for convolution kernel training, makes the image recognition of the final result is not so ideal. Although based on supervised learning The linear discriminant analysis (Linear Discriminant, Analysis, LDA) for training the convolution kernel deep learning model LDANet was proposed, but due to the limitation of the LDA algorithm's positive and negative samples separation ability, makes the LDANet model image recognition results there is no significant improvement. And, because of its unique sampling methods, make it easy over fitting phenomenon. The training effect is not satisfactory. In this paper, the PCANet and LDANet for the problem, mainly made the following contributions: 1) in this paper, the basic structure of PCANet based on Marginal Fisher Analysis (creative MFA) the introduction of convolution kernel training, proposes a new simple deep learning the architecture of MFANet. MFA take the supervised learning method, mapping and through the extracted feature vector, which is between the sample after projection distance as small as possible, at the same time. The distance between the card negative samples as large as possible. The heterogeneous characteristics and image features of similar images obtained by convolution convolution can be effectively separated in the new feature space. By using the standard data set: ICDAR2003 character recognition data set, the recognition performance of the proposed depth model test, and through the comparison of extraction model and image features some of the popular image classification, the final results show that the feature extraction ability of MFANet model is better than the other depth model.2) this paper proposes a new sampling method: the probability of random block level histogram based, to solve the robustness of PCANet model based on the core is not strong. The method is to calculate the pixel block probability, and based on the sampling probability to decide whether the pixel block features retained, finally using histogram output characteristics. Through the use of image recognition Other data sets: PIE FACE dataset, comparing the proposed classification results with the other popular classification results, and the experimental results show that the proposed method effectively reduces the over fitting phenomenon.

【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】