基于特征点群相似度计算模型的图像表示方法

针对空间金字塔匹配模型缺乏对图像中视觉物体旋转、平移和缩放的考虑问题,提出了一种基于特征点群相似度计算模型的图像表示方法.基于词汇树模型的粗匹配结果,通过特征点群拓扑、方向、距离等计算其相似度,并以此作为评价指标对匹配结果进行过滤;根据由特征点群计算所得的标准差椭圆的圆心、旋转角度对金字塔匹配的图像划分子区域并进行调整,从而得到图像抗旋转、平移和缩放的表示.分别在自建校园建筑物数据集和自建物体图像数据集上对方法进行了验证和比较,结果表明,该方法提高了分类识别的准确率和检索的查全率,特别是对于包含明显旋转、平移和缩放变化的图像数据效果更好.

An image representation method based on the similarity of feature points

To overcome the shortcoming of the Spatial Pyramid Matching (SPM) approach, which lacks invariance to translation, scale and rotation of visual objects in images, this paper proposes an image representation method based on the similarity of feature points. Firstly, it filters the rough matching result of bag-of-words by some properties including the topological similarity, the directional similarity and the distance similarity. Then, it adjusts the division of the image sub-regions according to the standard deviation ellipse center and the rotation angle of the feature points. Finally, the representation of anti-rotation, anti-translation and anti-scaling of image can be obtained. Experiments have been conducted by applying the proposed method to the campus building dataset and the object image dataset. It indicates that our method significantly improves the classification accuracy and recall ratio, especially for the dataset containing images with obvious rotation, translation and scaling transforms.