8-图像检索中的相关反馈技术方案.ppt

调整特征内部的权值 是M’维的rijk序列的标准方差值 归一化： 合理性？ 其它方法？ 小结 颜色 纹理 颜色 纹理 颜色 纹理 5.其他相关反馈 贝叶斯相关反馈技术 基于贝叶斯推理的图像检索 寻找最优映射g,使得分类错误的概率 降到最小。 最优映射g,由贝叶斯分类器定义： 贝叶斯相关反馈技术 所起的作用是提供第i类图像与查询相关的先验概率。 非常直观，将t-1刻系统计算得到的图像类符合用户查询的置信度作为t 时刻两者匹配的先验概率。 计算复杂度低。 Bayesian relevance feedback for content-based image retrieval The basic idea behind our proposal is the local estimation of the decision boundary between the “relevant” and “non-relevant” images in the neighborhood of the original query. The new query is then placed at a suitable distance from such boundary, on the side of the region containing relevant images. Problem Formulation The boundary of the region containing the images that are relevant to user query Q0 is depicted by the dotted line. The initial query Q0 and the neighborhood N(Q0) related to the k-nn search (k =5) are depicted. A new query computed in the mR–mN direction is shown such that its neighborhood (dashed line) is contained in the relevant region. Solution Assume a Gaussian model for the distributions of relevant and non-relevant images. Estimate the decision boundary between relevant and non-relevant images. Sub-optimal simplied models are often more effective than optimal complex models for estimating local decision boundaries in small sample size cases. Solution According to the Bayesian decision theory, such an approximation let us model the decision surface between these two “classes” of images as a hyperplane orthogonal to the line linking the means and passing through a point x0 defined by the following equation: Thus, the “optimal” query point should be located at a suitable distance from x0 in the (mR–mN ) direction. (i) (ii) the average variance of relevant and non-relevant images remains the same (iii) the “decision boundary” between relevant and non relevant images estimated according to the model illustrated in the previous section coincides with the decision boundary computed using the original neig