多特征融合的纹理图像分类研究

Research of Texture Image Classification Based on Multi-feature Fusion

为了提升纹理图像分类准确度,从图像单一特征出发,例如LBP特征、灰度共生矩阵特征、Gabor特征。首先对图像进行预处理,提取特征,运用机器学习方法进行分类,验证特征与分类器的有效性。再通过参数调整,最大程度上提高各方法在各特征基础上的准确率。单一特征包含图像某方面信息的侧重性,各特征有不同的关注点。为了克服特征的局限性,增加分类特征包含图像信息的全面性,提出一种特征融合的方法。在提取的纹理图像单一特征基础上,给各个特征设置相应权重,进行融合再分类。在brodatz纹理库图像上进行实验,得到由融合机制进行分类的准确度优于单一特征分类的结果。

In order to improve the classification accuracy of texture images,based on the single feature of images,such as lbp,gray co-occurrence matrix and gabor feature.Firstly,images are preprocessed,then features are extracted and texture images are classified by machine learning method,at last verify the validity of the features and classifiers.Through adjusting parameters,improve the accuracy in the maximum extent by each method on each feature.The Single feature contains particular information of image,which has different concerns from others.A strategy based on Multi-feature Fusion is proposed,which overcome the limitations of features and increase the comprehensiveness in images’information of the features used on classification.On the basis of extracting the single feature of texture images,the corresponding weights are set for each feature,and then images are classified by fusion feature.Experiments are conducted on the Brodatz texture image.The experimental results show the classification accuracy by feature fusion is better than results on single feature.