PHOCNet: A Deep Convolutional Neural Network for Word Spotting in Handwritten Documents

Abstract

In recent years, deep convolutional neural networks have achieved state of the art performance in various computer vision task such as classification, detection or segmentation. Due to their outstanding performance, CNNs are more and more used in the field of document image analysis as well. In this work, we present a CNN architecture that is trained with the recently proposed PHOC representation. We show empirically that our CNN architecture is able to outperform state of the art results for various word spotting benchmarks while exhibiting short training and test times.

Transferable Deep Features for Keyword Spotting

“…The core of the proposed method consists of using a CNN as a feature extractor. We have used PHOCnet [2], a CNN architecture recently proposed for segmentation-based KWS. PHOCnet was the best performing model on the recent ICFHR 2016 KWS competition (unpenalized MAP scenario) [1].…”

“…Regarding further details on the network architecture, as well as details on how training is performed (parameters, number of iterations, use of dropout, etc. ), the reader is referred to the original publication [2]. All layers between the input layer and the SPP layer are of variable size, as they depend on the input word image size.…”

“…Not surprisingly, learning-based methods are in general better performing than learning-free KWS methods [1]. The vast majority of recently proposed learning-based methods includes deep learning-based methods, which seem to have dominated this field as well [1][2][3].…”

“…As an alternative to the Euclidean distance, the Bray-Curtis dissimilarity (BC) has recently been employed in the context of keyword spotting [2,9]. We compare BC with the Euclidean distance, with and without applying L2-normalization before evaluation.…”

“…In this work, we use PHOCnet [2] as our model of reference. PHOCnet is a Deep Convolutional Network that has recently been proposed for KWS.…”