Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring Architectures for CNN-Based Word Spotting

Eugen Rusakov, Sebastian Sudholt, Fabian Wolf, Gernot A. Fink

TL;DR

The recently successful TPP-PHOCNet will be compared to a Residual Network, a Densely Connected Convolutional Network and a LeNet architecture empirically and it is shown that a complex model can be beneficial for word spotting on harder tasks but gives no advantage for easier benchmarks such as the George Washington Database.

Abstract

The goal in word spotting is to retrieve parts of document images which are relevant with respect to a certain user-defined query. The recent past has seen attribute-based Convolutional Neural Networks take over this field of research. As is common for other fields of computer vision, the CNNs used for this task are already considerably deep. The question that arises, however, is: How complex does a CNN have to be for word spotting? Are increasingly deeper models giving increasingly better results or does performance behave asymptotically for these architectures? On the other hand, can similar results be obtained with a much smaller CNN? The goal of this paper is to give an answer to these questions. Therefore, the recently successful TPP-PHOCNet will be compared to a Residual Network, a Densely Connected Convolutional Network and a LeNet architecture empirically. As will be seen in the evaluation, a complex model can be beneficial for word spotting on harder tasks such as the IAM Offline Database but gives no advantage for easier benchmarks such as the George Washington Database.

Exploring Architectures for CNN-Based Word Spotting

TL;DR

The recently successful TPP-PHOCNet will be compared to a Residual Network, a Densely Connected Convolutional Network and a LeNet architecture empirically and it is shown that a complex model can be beneficial for word spotting on harder tasks but gives no advantage for easier benchmarks such as the George Washington Database.

Abstract

The goal in word spotting is to retrieve parts of document images which are relevant with respect to a certain user-defined query. The recent past has seen attribute-based Convolutional Neural Networks take over this field of research. As is common for other fields of computer vision, the CNNs used for this task are already considerably deep. The question that arises, however, is: How complex does a CNN have to be for word spotting? Are increasingly deeper models giving increasingly better results or does performance behave asymptotically for these architectures? On the other hand, can similar results be obtained with a much smaller CNN? The goal of this paper is to give an answer to these questions. Therefore, the recently successful TPP-PHOCNet will be compared to a Residual Network, a Densely Connected Convolutional Network and a LeNet architecture empirically. As will be seen in the evaluation, a complex model can be beneficial for word spotting on harder tasks such as the IAM Offline Database but gives no advantage for easier benchmarks such as the George Washington Database.

Paper Structure

This paper contains 17 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Word Spotting
  4. Deep Learning
  5. ResNet
  6. DenseNets
  7. Architectures for Word Spotting
  8. PHOCLeNet
  9. TPP-PHOCNet
  10. PHOCResNet
  11. PHOCDenseNet
  12. Experiments
  13. Datasets
  14. Evaluation Protocol
  15. Training Setup
  16. ...and 2 more sections

Figures (1)

  • Figure 1: Overview of the explored CNN architectures. The networks are from left to right, PHOCLeNet, TPP-PHOCNet, PHOCResNet, and PHOCDenseNet. The CNNs are trained to predict a desired word string embedding generated from the annotation. Word spotting can then be performed in the embedding space through a simple nearest neighbor search using the cosine similarity as distance metric. Please note, that due to space issues this overview is not an exactly architecture description especially for the residual bottleneck and dense blocks. The exactly description is given in Sec. \ref{['sec:method']}.