基于字典优化的迁移稀疏编码方法

Dictionary Refinement-based Transfer Sparse Coding Method

传统的编码方法通常对字典采取随机初始化,极大影响图像分类精度。基于此,提出一种基于kmeans的字典优化方法,并将其与迁移稀疏编码相结合。先将图像中每个局部描述子投影到线性子空间,在此空间取距离特征最近的k个特征作为过完备字典,均衡的选择基向量来表达图像;同时考虑了图像的分布差异和局部特征,有效保证编码的稳定性。在三个跨域图像数据集上实验表明,与同类方法相比,该方法能显著提高跨域分类性能。

The traditional feature coding method usually adopts dictionary initialization randomly,which greatly affects classification accuracy.Based on this,a dictionary refinement-based K-means clustering method combined with transfer sparse coding have been proposed.First,each local description subspace in the image is projected into the local linear subspace,in which we choose K features closest to the features are taken as over complete dictionary.those dictionary can steady express the image.what’s more,we also introduce the distribution differences and local features of the image to objective function,which effectively ensure the stability of coding.Experimental results on three cross domain image datasets show that the proposed method can significantly improve the performance of classification compared with similar methods.