[Paper Review] Expolring Architectures for CNN-Based Word Spotting.

Eugen Rusakov, Sebastian Sudholt+2 more|arXiv (Cornell University)|Jun 28, 2018

Handwritten Text Recognition Techniques23 references3 citations

TL;DR

This paper evaluates the impact of CNN architecture depth on word spotting by comparing TPP-HOCNet against ResNet, DenseNet, and LeNet on the IAM and George Washington databases. Results show deeper models improve performance on hard tasks like IAM but offer no benefit on easier benchmarks such as George Washington, indicating performance asymptotes with increased complexity for simpler tasks.

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 bet- ter 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.

Motivation & Objective

• To investigate whether increasingly deeper CNN architectures improve word spotting performance.
• To determine if simpler, smaller CNNs can achieve comparable results to complex models.
• To evaluate the trade-off between model complexity and performance on different word spotting benchmarks.
• To compare the effectiveness of state-of-the-art architectures like ResNet, DenseNet, and LeNet against the recent TPP-HOCNet.
• To identify whether performance gains plateau with increased depth for word spotting tasks.

Proposed method

• The study employs four CNN architectures: TPP-HOCNet, ResNet, DenseNet, and LeNet for word spotting.
• Each model is trained and evaluated on two benchmark datasets: the IAM Offline Database and the George Washington Database.
• Word spotting performance is measured using standard retrieval metrics, such as mean average precision (mAP).
• The comparison focuses on how architectural depth and connectivity patterns affect retrieval accuracy.
• Empirical evaluation is conducted across varying difficulty levels of document image datasets.
• The analysis includes ablation-like comparisons to assess the contribution of model depth and structure.

Experimental results

Research questions

• RQ1 Does increasing CNN depth consistently improve word spotting performance across all datasets?
• RQ2 Can a smaller, shallower CNN achieve performance comparable to deeper architectures?
• RQ3 Is there a performance saturation point for deeper models in word spotting tasks?
• RQ4 How do different architectural designs (e.g., residual, dense, plain) compare on word spotting benchmarks?
• RQ5 Does model complexity provide a significant advantage on easier versus harder word spotting datasets?

Key findings

• On the IAM Offline Database, a complex model like TPP-HOCNet provides a performance advantage over simpler architectures.
• On the George Washington Database, deeper models such as ResNet and DenseNet do not outperform simpler models like LeNet.
• Performance gains from increased model depth plateau on easier benchmarks, indicating asymptotic behavior.
• The study finds that architectural complexity is beneficial only for challenging word spotting tasks with high variability.
• Smaller networks like LeNet achieve competitive results on simpler datasets, suggesting that deeper models are not always necessary.
• The results suggest that model choice should be guided by dataset difficulty rather than depth alone.