首先将投影方法运用于图像局部子块,并与矢量量化技术相结合,得到投影矢量索引直方图特征,能够有效提取图像的颜色分布、空间关系等信息;其次,提出基于块基元的LBP纹理特征算法,既能有效提取块基元的结构模式,又能避免传统基于单像素LBP模板的不稳定性,并大幅减少了计算量;最后,基于显著图提出了划分显著区域和非显著区域的特征提取方案,使得分别提取的特征更富于视觉意义。实验结果表明,本文算法相对于传统索引直方图方法在性能上有较大的提高,平均查准率平均提高幅度为6.39%。

A projection method is applied to local image blocks through combinations with Vector Quantization(VQ) to generate the projection vector index histogram, which can efficiently represent the distribution and the spacial information of colors. Furthermore, a block-based local binary pattern(LBP) algorithm is proposed, which can effectively extract the structural model of block primitive, avoid instability,and reduce computations compared to the traditional methods. Finally, an image is partitioned into significant regions and non-significant regions based on its saliency map to give the features which are extracted from them more visual sense. Our proposed method is improving the performance by an average of 6.39% compared to the classical index histogram algorithm.