一种基于最优路径搜索的图像分类方法

Image Classification Algorithm Based on Optimal Path Search

目前已经有多种基于树的算法来解决多类别图像分类问题,然而由于选择的学习和贪婪预测策略不当,这些算法在分类精度和测试时间效率间不能实现很好的均衡。提出一种新的分类器,当树形架构已知时能在效率和精度间实现很好的折中。首先,将图像分类问题转化为树结构中最优路径的搜索问题,提出新的类似于分支界定的算法来实现最优路径的高效搜索。其次,使用结构化支持向量机(SSVM)在多种边界约束下联合训练分类器。仿真实验结果表明,相对于当前最新"基于树"的贪婪算法,当应用于Caltech-256、SUN和Image Net 1K等数据集时,该算法在效率较高时的精度分别上升了4.65%,5.43%和4.07%。

Many algorithms based on tree are proposed to solve the image classification problem for a large number of categories. Due to learning and greedy prediction strategy choice of undeserved, methods based on tree - based representations cannot aehieve good trade - off between accuraey and test time efficieney. In this paper, a elassifier is proposed which achieves a better trade - off between efficiency and aecuracy with a given tree - shaped hierarchy. Firstly, the image classification problem is eonverted as finding the best path in the tree hierarchy, and a novel branch and bound - like algorithm is introduced to efficiently search for the best path. Secondly, the elassifiers are trained using a Structured SVM （SSVM） formulation with various bound constraints. Simulation results show that, this method achieves a significant 4. 65% , 5.43%, and 4. 07% improvement in aceuracy at high effieieney compared to state -of- the -art greedy ＂tree- based＂ methods on Caltech- 256, SUN and ImagcNet 1K dataset, respectively.