多尺度显著性区域提取的模板匹配

针对模板匹配过程中强遮挡、剧烈背景变化及物体非刚性形变等难题,本文提出了一种基于多尺度显著性区域提取的模板匹配算法。算法采用多尺度-显著性特征并行提取的方式:一方面利用空间金字塔模型将参考图像中的模板和待匹配图像中的目标区域分割成不同尺度的网格,采用可形变多相似性度量方法(Deformable Diversity Similarity,DDIS)计算不同尺度下的匹配得分;同时,算法提取模板区域的显著性区域图,形成模板区域的显著性得分;随后,利用显著性得分对不同尺度的匹配得分进行加权融合,在融合得到的匹配得分图上寻找最佳匹配区域。算法与取得目前最好结果的DDIS方法相比,AUC(Area Under Curve)指标提升2.9%。实验结果表明,显著性区域提取使匹配方法更加关注目标物体,削弱背景及遮挡物体对其影响,从而增强模板匹配方法对于背景变化及遮挡的抵抗能力。另外,空间金字塔模型能够增强模板匹配方法对于物体不同尺度下的特征提取,如物体的局部轮廓及结构特征等。二者结合有效地提高了匹配精度。

Template matching with multi-scale saliency

Traditional template matching methods suffer from heavy occlusion, intense background change and non-rigid deformation. A multi-scale saliency template matching method is proposed in this article in order to deal with such conditions. The method extracted saliency and multi-scale features in parallel. On the one hand, the template and the target images were first divided into grids of different scales using spatial pyramid model. Deformable Diversity Similarity (DDIS) was calculated under such different grids. On the other hand, saliency map of the template image was calculated using saliency segmentation method. Such saliency map s are then used to weight the scores calculated by DDIS under different grids. Finally, the final score map is calculated by fusing the score maps under different grids. The method proposed achieves 2.9% AUC(Area Under Curve) improvement compared with original DDIS method. Experiments show that salient object segmentation helps the method to focus more on object than background, therefore improve the robustness to background changes and occlusion. Besides, spatial pyramid model makes the method to consider information from different scale, for example, local contours and structural features of an object. Combining these two factors raises the matching accuracy significantly.