自然场景图像与合成图像的快速分类

发布时间：2018-01-19 01:35

本文关键词： 图像类型快速分类特征提取词袋模型层次化分类算法　出处：《中国图象图形学报》2017年05期 　论文类型：期刊论文

【摘要】：目的随着现代通信和传感技术的快速发展,互联网上多媒体数据日益增长,既为人们生活提供了便利,又给信息有效利用提出了挑战。为充分挖掘网络图像中蕴含的丰富信息,同时考虑到网络中图像类型的多样性,以及不同类型的图像需要不同的处理方法,本文针对当今互联网中两种主要的图像类型:自然场景图像与合成图像,设计层次化的快速分类算法。方法该算法包括两层,第1层利用两类图像在颜色,饱和度以及边缘对比度上表现出来的差异性提取全局特征,并结合支持向量机(SVM)进行初步分类,第1层分类结果中低置信度的图像会被送到第2层中。在第2层中,系统基于词袋模型(bag-of-words)对图像不同类型的局部区域的纹理信息进行编码得到局部特征并结合第2个SVM分类器完成最终分类。针对层次化分类框架,文中还提出两种策略对两个分类器进行融合,分别为分类器结果融合与全局+局部特征融合。为测试算法的实用性,同时收集并发布了一个包含超过30 000幅图像的数据库。结果本文设计的全局与局部特征对两类图像具有较强的判别性。在单核Intel Xeon(R)(2.50 GHz)CPU上,分类精度可达到98.26%,分类速度超过40帧/s。另外通过与基于卷积神经网络的方法进行对比实验可发现,本文提出的算法在性能上与浅层网络相当,但消耗更少的计算资源。结论本文基于自然场景图像与合成图像在颜色、饱和度、边缘对比度以及局部纹理上的差异,设计并提取快速有效的全局与局部特征,并结合层次化的分类框架,完成对两类图像的快速分类任务,该算法兼顾分类精度与分类速度,可应用于对实时性要求较高的图像检索与数据信息挖掘等实际项目中。
[Abstract]:With the rapid development of modern communication and sensing technology, multimedia data on the Internet is increasing day by day, which provides convenience for people's life. In order to fully mine the abundant information contained in the network image, consider the diversity of the image type in the network, and different image need different processing methods. In this paper, we design a hierarchical fast classification algorithm for two main image types in the Internet: natural scene image and synthetic image. The algorithm includes two layers, and the first layer uses two kinds of images in color. The differences in saturation and edge contrast are used to extract the global features, and the support vector machine (SVM) is used to classify the features. Images of low confidence in the results of the first level of classification are sent to layer 2. In layer 2. The system is based on bag-of-words-based model. The local features are obtained by encoding the texture information of different types of local regions of the image and the final classification is completed with the second SVM classifier. Two strategies are also proposed to fuse the two classifiers, one is the fusion of the results of the classifier and the other is the fusion of the global local features, and the other is the practicability of the test algorithm. At the same time, a database containing more than 30,000 images was collected and published. Results the global and local features designed in this paper have strong discriminant for the two kinds of images. 2. 50 GHz)CPU. The classification accuracy can reach 98.26 and the classification speed is more than 40 frames / s. In addition, it can be found by comparing with the method based on convolution neural network. The proposed algorithm has the same performance as the shallow network, but consumes less computing resources. Conclusion based on the natural scene image and the composite image in color, saturation. Edge contrast and local texture differences, design and extract fast and effective global and local features, and combine the hierarchical classification framework to complete the task of fast classification of the two kinds of images. This algorithm can be applied to image retrieval and data information mining with high real-time requirement, because of the combination of classification accuracy and classification speed.
【作者单位】：哈尔滨理工大学自动化学院;中国科学院自动化研究所模式识别国家重点实验室;
【基金】：国家自然科学基金项目(61411136002)~~
【分类号】：TP391.41
【正文快照】： 0引言随着互联网、智能手机和通信技术的迅速发展,互联网上的图像数据在快速增长。海量的图像可在网络信息分析、商业数据挖掘、敏感信息检测等领域中发挥出巨大作用,因此,网络图像的处理、分析、搜索和理解得到了学术界和工业界的广泛重视。考虑到网络中图像类型的多样性,以

【相似文献】