Handwriting Recognition (HR) of Family History Documents using a 2-D Warping-based Word-level HR Approach

An enormous amount of handwritten information exists that is potentially very useful for family history research. However, finding information of interest is a daunting task unless the handwriting is transcribed or indexed so that it can be digitally searched. Transcription / indexing is typically done manually because automatic handwriting recognition (HR) is not yet accurate enough to provide reliable transcriptions. Since manual transcription is both costly and time consuming, improvements in HR are very desirable. In this paper, we describe a novel method of word-level HR that we recently published at the International Conference on Document Analysis and Recognition (ICDAR 2011) and discuss how it can be applied to family history document images. We use an automatic morphing algorithm to generate a 2-D geometric warp that aligns each unknown word to known training examples. Once the word strokes are aligned, a distance map is used to calculate how different the aligned (warped) word is from the training example. The label of the training example that is most similar is used as the digital transcription for the previously unknown word. Our initial results are based on two datasets, each consisting of 1,000 training words and 1,000 test words. For in-vocabulary words, we get 88.77% and 89.33% word recognition accuracy, respectively.