基于支持向量机和半监督深度信念网络学习的图像分类算法研究

发布时间：2018-03-20 00:19

本文选题：图像分类　切入点：支持向量机　出处：《江西理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：图像分类是计算机视觉领域的研究热点之一,随着网络资源的日益多样性,图像规模愈来愈大,内容愈来愈复杂,表现形式也愈来愈多样,图像的复杂化使得图像分类技术面临着巨大的挑战。本文综合目前图像分类的相关算法,概括出现存算法的不足之处并给出相应的对应方法。针对小规模图像分类技术的单一性与局限性、现存的大规模图像分类精度低的问题,分别给出了不同的解决方案并进行了详细的阐述。为探求性能更高的小规模分类图像技术,提出基于Daubechies小波的快速PCA和SVM(DW-FPSVM)的图像分类算法。先对图像进行一次二维Daubechies小波分解并提取人脸特征,接着利用快速PCA对特征降维去噪处理,最后建立支持向量机模型对处理好的特征进行分类。本文以ORL人脸图像库作为输入展开实验,得出:图像分类的准确率随支持向量机核参数?的增大而减小且DW-FPSVM算法能够有效的提高图像分类的准确性和稳固性。除此之外,还从训练时间和分类时间将DW-FPSVM与其它算法的相关时间做了比较,证明了该算法的高效性。针对浅层次大规模图像分类的低精度问题,提出深层次特征学习的Adaboost图像分类算法(AICDFL)。首先以DBN作为弱分类器对样本图像进行学习,根据每次训练得到的错误率以及各样本的分类准确性调整权值;然后使用BP算子回溯再次整体调整体样本权值并输出调整后的每个分类器的错误率,最后将所有弱分类器集成强分类器,输出最终结果。本文使用MNIST和ETH-80两种数据集进行实验仿真,并将分类结果与其他算法的分类结果进行比较。结果表明AICDFL算法最优,有效实现高精度的大规模图像分类,更具应用优势。
[Abstract]:Image classification is one of the research hotspots in the field of computer vision, with the growing diversity of cyber source, the image scale is more and more big, the content is more complex, form more and more diverse and complicated image so that the image classification technology is facing a huge challenge. The comprehensive algorithm for image classification at present, sums up deficiencies of the algorithm and the corresponding method. The single and the limitations of small scale image classification technology, large-scale image classification accuracy of the existing problem of low, different solutions are given and described in detail. In order to seek higher performance of small scale image classification technology, proposed Daubechies wavelet fast PCA and SVM (DW-FPSVM) based on the image classification algorithm. First the image of a two-dimensional Daubechies wavelet decomposition and feature extraction, and then use the fast PCA For feature reduction denoising, finally establish the model of support vector machine to classify the deal. Based on the ORL face image database as input expansion experiments that the accuracy of image classification with SVM kernel parameters? Decreases and the DW-FPSVM algorithm to image classification accuracy and improve the stability in addition, also makes a comparison between the DW-FPSVM algorithm and other related time from the training time and classification time, show the efficiency of the proposed algorithm. Aiming at the problem of low precision of shallow level large-scale image classification, proposed Adaboost image classification algorithm study deep level feature (AICDFL). Using DBN as a weak classifier study on the sample image, according to the error rates of each training sample and the classification accuracy of adjusting weight; and then use the BP operator back again to adjust the overall body like the weight The error rate and the output of each classifier is adjusted, finally all the weak classifiers strong classifier, the results of the final output. This paper uses MNIST and ETH-80 two data set for experiment simulation, classification and the classification results were compared with the results of other algorithms. The results show that AICDFL algorithm is optimal, effective implementation of large-scale image classification with high accuracy and more advantages.

【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】