A multiple instances approach to improving keyword spotting on historical Mongolian document images

Keyword spotting in handwritten document images using supervised and unsupervised representations

“…The authors employ the Boostmap algorithm described in [259] to embed the feature space of variable-length representations which are matched with DTW into a Euclidean space for faster comparisons. In the same direction, Wei et al [118] use DFT on variable-length word profiles to create fixed-length vectors.…”

“…The authors in [118] present five scored-based and three rank-based fusion methods to merge multiple ranked lists obtained from each top-ranked instance on the initial ranking list. Since the similarity scores among separate ranked lists may differ both in range and distribution, they also suggest to normalize these scores using a number of score normalization techniques.…”

“…By these means, the system is evaluated in terms of how accurately the query instance is retrieved. We should also note here that in a segmentation-free method under the QBE scenario, the query itself should be taken into account in the final hit list, since it could be missing from Word Yes Multi-writer Precision/Recall, F-measure [106] Word-part Yes Multi-writer Precision/Recall rates [124] Word No Printed Mean Precision/Recall [195] Word Yes Single Precision/Recall [237] Free Yes Multi-writer MAP [241] Line No Printed MAP [59] Free Yes Single MAP [118] Word No Multi-writer R-Precision [102] Free No Multi-writer R-Precision [153] Word Yes Single NDCG [167] Word Yes Single Soft(edit distance) MAP the retrieved regions.…”