基于部分连接神经网络的场景识别  

Scene Recognition Based on Partially Connected Neural Network

作　　者：张月[1] 潘伟[1] 陈晓熹[1] 

机构地区：[1]厦门大学信息科学与技术学院,福建省仿脑智能系统重点实验室,福建厦门361005

出　　处：《厦门大学学报（自然科学版）》2010年第4期482-486,共5页Journal of Xiamen University：Natural Science

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

摘　　要：目前基于图像的场景识别的方法都依赖于对图像特征的选取及特征数目的精简.提出了一种基于部分连接演化神经网络模型来进行图像场景识别的新方法:不对图像进行特征提取,而是将待识别图像的每个像素都作为神经网络的输入.为了克服新方法由于大量神经元引起的模型训练时间过长问题,将基于C语言计算架构的演化神经网络模型创造性地移植到基于图形处理器(GPU)的通用并行计算构架(CUDA),神经网络的演化训练速度提高200倍以上.在实验中,尽管输入的图像大小达到300×400像素(120 000个输入神经元),但CUDA的部分连接演化神经网络对场景图像有较强的识别能力,对亮度、缩放、旋转等变化也有较好的鲁棒性.At precent, the method of scene recognition which is based on images is dependent on the selection of images features and the number of characteristics. This paper presents a partially connected evolutionary neural network model to recognite seene： It doesn＇t extract feature from the images, but make each pixel as the input of neural network. In order to overcome the problem of the new method that the time of model training was too long, which caused by large number of neurons. In this paper,we put the partial- ly connected evolutionary neural network which based on C language computing architecture to CUDA computing architecture, the e- volution training of neural network had improved 200 times. In the experiment, although the input image size are 300 × 400 pixels （120 000 input neurons）, but partially connected evolutionary neural network which based on CUDA architecture has not only a strong recognition ability in scene recognition, but also has good robustness against image transformation such as illumination,rotation and scale transformation.

关 键 词：场景识别 部分连接演化神经网络 CUDA 

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]