Hierarchical representation learning using spherical k-means for segmentation-free word spotting

Automatic segmentation-free and training-free word spotting is a challenging task due to the large intra-class variability of handwritten shapes and the need to process the whole document image. In this work, a novel unsupervised hierarchical handwriting representation is introduced, where the spherical k-means algorithm is used to learn a hierarchy of features for representing document images. A matching system is then employed for word spotting, which consists of two stages: (1) a fast pre-selection stage applying a sliding-window approach over compressed document image representations, and (2) a re-ranking stage based on a discriminative description that encodes the spatial layout of local features. The proposed approach is evaluated using three well-known benchmark datasets, the Lord Byron (LB), the George Washington (GW) and the IAM datasets. Results show our method to yield competitive performance compared to state-of-the-art approaches for segmentation-free and training-free word spotting. In addition, since our proposed framework has a low computational and memory complexity, it can be applied to large datasets.