一种无特征提取的自然场景图像分类新方法  

A New Non-feature Extraction Method for the Classification of Nature Scene Image

作　　者：王丽萍[1] 潘伟[1] 

机构地区：[1]厦门大学信息科学与技术学院,福建厦门361005

出　　处：《厦门大学学报（自然科学版）》2011年第4期697-702,共6页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金项目(60975084);福建省自然科学基金项目(2009J01305)

摘　　要：提出了一种新的基于部分连接神经网络的自然场景图像分类方法.运用该方法对图像进行模式识别时,不必进行特征提取,而是将整个图像输入神经网络,由神经网络在训练中透明地选择和识别特征.由于大型图形处理器(GPU)并行处理系统的运用,使得神经网络演化速度大大加快,弥补了该方法计算量大的弱点.实验结果表明,利用部分连接神经网络进行场景图像分类,与利用特征提取后再识别场景的分类方法比较,在总识别率上大体相当;但不必进行特征提取,而且速度很快.并且,还运用了插值和延拓两种方法来对图像进行尺寸调整,使得神经网络可以训练和识别不同大小的场景图像.This paper presented a novel method for scene images classification via partially connected neuron network.The Parcone module didn′t need any feature extraction in the process of scene classication.In the training process of the neuron network,we inputed the entire image to the Parcone and the module could select the recognition feature automatically.The using of large-scale GPU parallel computing system accelerated the evolutionary of the neuron network in training and overcame the weakness resulted from the large calculated amount.The experiments showed that the Parcone performed as equally well as other methods using feature extraction in the average recognition rate in scene classification.Moreover,the method using Parcone was more effective in the computation speed.Besides,we used two methods of interpolation and continuation to adjust the sizes of the scene images so that the Parcone could recognize the scene images of different sizes.

关 键 词：场景分类 部分连接神经网络 插值 延拓 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程]