结合空间上下文的局部约束线性特征编码

摘要

针对传统特征编码方法聚焦于在特征空间进行编码,忽略了图像内容的空间信息,导致图像表达不准确、分类精度较低的问题,提出一种在特征空间中以图像空间上下文信息为导向的局部特征编码方法.首先基于最近邻原则为每个局部特征点选择字典中心作为向量基;然后采用探测局部特征的相邻特征点方法建立图像空间上下文约束,并将其用于特征相似性判别;再根据预设阈值来更新向量基,将其用于重构特征;最后将图像的稀疏向量用于分类器进行图像分类.实验结果表明,与同类方法相比,该方法能显著地提高分类精度,更利于图像分类.%Existing traditional coding methods mostly focus on the feature space regardless of the spatial domain of the image,which can cause deviation of the image expression and misclassification.In order to improve accuracy of category,feature coding both in feature space and spatial domain of the image will be taken into account while preserving locality.Firstly,we choose some nearest words as basis vector based on feature space and introduce spatial context to measure descriptors' similarity.Then,we can consider whether to update basis vector according to the threshold and encode the local features with locality-constrained coding.Finally,we get final sparse vector of image and train classifier model to category.Compared with other methods,the experimental results shows that our method can improve classification accuracy and be beneficial for classification.