Word Image Representation Based on Sequence to Sequence Model with Attention Mechanism for Out-of-Vocabulary Keyword Spotting

WSRNet: Joint Spotting and Recognition of Handwritten Words

“…The sequence-to-sequence architecture has led to state-of-the-art results in Natural Language Processing, involving translating an input sequence to an output sequence of a different length in general. Use of the Seq2Seq architecture has started been used in HTR and KWS recently as well [26,27].…”

Keyword spotting in handwritten document images using supervised and unsupervised representations

“…With the recent advent of deep learning-based KWS methods, a standard solution for example model architectures is to normalize input images to fixed size [69,104,121,166]. For instance, Wei and co-workers propose a normalization by resizing all input images to a standard size of 310 pixel width and 50 height in [121], whereas in [120], they resize all input images so that they have the same width (either pure or by padding white pixels) and aspect ratio. Wicht et al [150,151] normalize the word images to remove the skew and slant of the text using [228].…”

“…The middle zone is modeled using HMM whereas the upper/lower zones are used to train similar feed-forward networks that include convolutional layers in their architecture have been used. These networks work typically either by producing in their output a suitable descriptor of the input word image [4,100,101,159,173,174], or by using network layer activations to create input word image descriptors [32,34,37,74,120,153,180,235]. Again, a typical distance that is used is the Euclidean.…”

“…Once the string embedding model is trained it is used to teach an image embedding model [266] so that, given a word image as input, it regresses at its Finally, Retsinas et al [37] propose a unified model that can handle both KWS and word recognition with the same network architecture. The network is comprised of a non-recurrent CTC branch and a sequence-to-sequence (Seq2Seq) branch [120] which is used to create efficient word representations. The deep features are further augmented with an autoencoding module which translates query strings to the Seq2Seq intermediate representation space, or by forced aligning the query to the decoder.…”