基于感受野学习的特征词袋模型简化算法

本文研究了在图像识别任务中，感受野学习对于特征词袋模型的影响。在特征词袋模型中，一个特征的感受野主要取决于视觉词典中的视觉单词和池化过程中所使用的区域。视觉单词决定了特征的选择性，池化区域则影响特征的局部性。文中提出了一种改进的感受野学习算法，用于寻找针对具体的图像识别任务最具有效性的感受野，同时考虑到了视觉单词数量增长所带来的冗余问题。通过学习，低效、冗余的视觉单词和池化区域会被发现，并从特征词袋模型中移除，从而产生一个针对具体分类任务更精简的、更具可分性的图像表达。最后，通过实验显示了该算法的有效性，学习到的模型除了结构精简，在识别精度上相比原有方法也能有一定提升。
In this work, the effects of receptive field learning on a bag-of-features pipeline were investigated for an image identification task. In a bag-of-features model, the receptive field of a feature depends mostly on use of visual words in a visual dictionary and the region used during the pooling process. Codewords make the feature respond to specific image patches and the pooling regions determine the spatial scope of the features. A modified graft feature selecting algorithm was proposed to find the most efficient receptive fields for identification purposes; this considers the redundancy problem created by simultaneously increasing visual words. Using learning receptive fields, inefficient and redundant codewords and pooling regions were found and subsequently eliminated from the pooling region, this made the pipeline more compact and separable for the specified classification task. The experiments show that the modified learning algorithm is effective and the learned pipeline useful for both structural simplification and improving classification accuracy compared with the baseline method