基于压缩感知理论的RGB-D图像分类方法  

RGB-D IMAGES CLASSIFICATION BASED ON COMPRESSED SENSING THEORY

作　　者：黄晓琳[1] 薛月菊[2] 涂淑琴[2] 李鸿生[1] 何金辉[2] 

机构地区：[1]华南农业大学工程学院,广东广州510642 [2]华南农业大学信息学院,广东广州510642

出　　处：《计算机应用与软件》2014年第3期195-198,共4页Computer Applications and Software

基　　金：国家科技支撑计划课题(2013BAJ13B05)

摘　　要：三维图像分类能有效克服二维彩色图像分类易受光照变化、阴影、物体遮挡以及环境变化等因素的干扰。利用压缩感知的方法研究Kinect相机获取的带深度信息的RGB-D图像分类问题。该方法首先利用下采样和PCA的方法分别对RGB图像和深度图像进行特征提取;再将所提取的特征信息融合;然后利用压缩感知方法对融合后的特征信息进行稀疏分解并分类。最后,利用该方法对6类蔬菜、7类水果及文件夹和相机等共15类RGB-D图像进行分类实验,比较了压缩感知方法和SVM分类器的RGB-D图像分类精度,并对比分析了深度信息对图像分类精度的影响。实验表明,压缩感知方法对RGB-D图像分类精度高于SVM的精度,且加入深度信息的RGB-D图像分类精度高于RGB图像的精度。3D images classification can effectively overcome the disadvantage of 2D colour images classification such as susceptible to the interferences of illumination changes, shadows, objects occlusion, environmental changes and other factors. In this paper, compressed sensingis used to study the classification of RGB-D images containing deep information gained by Kinect camera. First, the method uses downsampling and PCA to extract the features from RGB images and depth images separately. Secondly, the extracted features are fused. And then the sparse decomposition of the fused features is conducted by using compressed sensing and followed by classification. Finally, this method is used to make classification experiments on images in RGB-D dataset, including 6 kinds of vegetables, 7 kinds of fruits, and the binder, the camera, counted 15 classes in total. The precision of RGB-D images classification using compressed sensing is compared with that of SVM classifier; and the impact of depth information on image classification precision is compared and analysed as well. Experiments show that the classification precision of RGB-D images is higher by using compressed sensing than using SVM, and the classification precision of RGB-D images with depth information added is higher than that of RGB images.

关 键 词：RGB—D图像分类 压缩感知 特征提取 稀疏分解 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]