Study on learning techniques which utilize existing classifiers

我们已经研究了利用仍然存在的分类器来构造新类别的分类器的方法。研究了相似图像搜索和图像分割问题的方法。对于图像搜索，给定一幅查询图像，通过实例学习得到合适的权向量。虽然该方法的搜索质量很好，但它是耗时的。为了减少计算成本，我们提出了一种方法，近似权向量从现有的分类器（权向量）。我们比较了基于降维的方法和基于PCA的方法的搜索质量，发现基于PCA的方法优于其他近似技术。对于图像分割，该任务被制定为能量最小化。我们尝试使用现有的图像分类器来改进能量函数。我们已经证明了分割质量可以通过限制可能的对象类别来提高。

We have investigated methods to utilize still existing classifiers to construct classifier for a novel category. We have studied the methods for similar image search and image segmentation problems. For image search, given a query image, appropriate weight vector is obtained via learning from examples. Although the search quality of the method is very good, it is time consuming. To reduce the computational cost, we proposed a method which approximates weight vectors from existing classifiers (weight vectors). We compared the search quality with the method based on the dimensionality reduction and found PCA-based method can outperform the other approximation techniques. For image segmentation, the task is formulated as the energy minimization. We tried to improve the energy function using the existing image classifier. We have shown the segmentation quality can be improved by limiting possible object categories.