为了准确地对不同学习样本数的图像进行分类,首先讨论了对属于不同类别的图像序列如何进行正确有监督分类的问题,由于解决这类问题首先要选取适合图像分类的图像特征作为分类的依据,为此先用图像角点来构成Delaunay图,然后将由Delaunay图的谱特征形成的分类特征矢量作为分类的依据;其次,由于分类器的选取也直接影响分类结果,为此采用了学习效率高的概率神经网络分类器来进行分类.经过大量分类实验表明,图谱特征很好地保持了图像的结构特征,是理想的图像分类特征;经过与其他相关分类器的分类比较实验表明,基于概率神经网络的分类器可以准确地进行图像分类;通过不同学习样本数的比较,证实了概率神经网络在进行图像分类时,对于学习样本数并不敏感,并具有一定稳定性.

This study applies graph spectra and probabilistic neural networks to the task of supervised images classification. At first, comers are extracted from images to construct Delaunay graphs. These relational graphs have proved alluring as structural representations for images. Then graph spectral decomposition method is adopted to get eigenvalues and eigenvectors of the adjacency matrix. Graph spectra can preserve the primary structure of graphs. The spectrum of these graphs will be used as feature vectors for classification. At last probabilistic neural networks will give the classification result according to the vectors . As for the classifier, PNN has high speed of learning because the learning rule is simple and new trainings patterns can be incorporated into a previously trained classifier quite easily, which might be important for a particular on-fine application. Experimental results show that this method can achieve the best result of images classification.