基于多类神经网络机的自然图像分类

On Applying Multi-Class Neural Network to Classifying Scene Image

基于底层视觉特征把图像分为具有特定意义的类别,对于基于内容的图像检索意义重大。因为在这种分类基础上,可以在图像库中建立一种有效的索引机制。在底层视觉特征方面,文中主要提取了图像的主颜色特征和GABOR纹理特征,然后,提出了一种多类神经网络机用于图像的分类。在一个含有4000幅的图像库中,实验结果证明这种方法可以达到70%以上的准确率。

Refs.2 and 3 provide two-class classifiers to classify scene images. We propose using multi-class neural network to do a better job of classifying scene images. Proper classification is helpful to retrieving satisfactorily content-based scene image. Our method requires extracting two low-level features: dominant color feature and Gabor texture feature. We extract only the dominant color because we want to save memory space and computation time. In the multi-class RBF (Radial Basis Function) neural network we designed and implemented for experimental research, we paid particular attention to the fusion of the two low-level features: dominant color and Gabor texture. We performed fairly extensive but limited experiments with our method and experimental setup. For 8-class scene images, we attain an accuracy of more than 70% correct classification, higher than the 40%～50% correct classification attainable with Bayes classifiers.