[Paper Review] Dense Residual Network: Enhancing Global Dense Feature Flow for Character Recognition

Zhao Zhang, Zemin Tang+5 more|arXiv (Cornell University)|Jan 23, 2020

Handwritten Text Recognition Techniques51 references4 citations

TL;DR

This paper proposes Fast Dense Residual Network (FDRN), a novel CNN architecture that enhances global dense feature flow for scene text recognition by introducing a fast residual dense block (f-RDB) for efficient local feature fusion and a global dense block (GDB) for adaptive, holistic learning of global residual features. FDRN achieves state-of-the-art performance on multiple text recognition benchmarks by effectively integrating hierarchical features across all layers with reduced computational cost.

ABSTRACT

Deep Convolutional Neural Networks (CNNs), such as Dense Convolutional Networks (DenseNet), have achieved great success for image representation by discovering deep hierarchical information. However, most existing networks simply stacks the convolutional layers and hence failing to fully discover local and global feature information among layers. In this paper, we mainly explore how to enhance the local and global dense feature flow by exploiting hierarchical features fully from all the convolution layers. Technically, we propose an efficient and effective CNN framework, i.e., Fast Dense Residual Network (FDRN), for text recognition. To construct FDRN, we propose a new fast residual dense block (f-RDB) to retain the ability of local feature fusion and local residual learning of original RDB, which can reduce the computing efforts at the same time. After fully learning local residual dense features, we utilize the sum operation and several f-RDBs to define a new block termed global dense block (GDB) by imitating the construction of dense blocks to learn global dense residual features adaptively in a holistic way. Finally, we use two convolution layers to construct a down-sampling block to reduce the global feature size and extract deeper features. Extensive simulations show that FDRN obtains the enhanced recognition results, compared with other related models.

Motivation & Objective

• To address the limitation of existing CNNs in fully exploiting hierarchical local and global features across all layers.
• To improve feature flow in deep networks by enhancing both local residual learning and global dense feature aggregation.
• To design a computationally efficient architecture that maintains high performance in scene text recognition tasks.
• To enable adaptive, holistic learning of global residual features through a novel block structure inspired by dense blocks.
• To reduce computational overhead while preserving or improving recognition accuracy through optimized residual and dense connections.

Proposed method

• Proposes a fast residual dense block (f-RDB) that retains local feature fusion and residual learning while reducing computational cost.
• Introduces a global dense block (GDB) that mimics dense block design to learn global residual features in a holistic, adaptive manner.
• Employs sum operations and multiple f-RDBs within the GDB to aggregate features across layers for enhanced global representation.
• Uses two convolutional layers as a down-sampling block to reduce feature map size and extract deeper representations.
• Combines f-RDBs and GDBs in a stacked, hierarchical fashion to progressively build deep, hierarchical features.
• Architects the full FDRN network by stacking f-RDBs, GDBs, and down-sampling blocks to optimize feature flow and representation.

Experimental results

Research questions

• RQ1 How can local and global feature flows be enhanced in deep CNNs for improved text recognition?
• RQ2 What is the impact of integrating local residual learning with global dense feature aggregation in a unified architecture?
• RQ3 Can a lightweight block design (f-RDB) maintain performance while reducing computational cost in text recognition networks?
• RQ4 How does the proposed global dense block (GDB) improve feature representation compared to standard residual or dense blocks?
• RQ5 To what extent does the hierarchical integration of features across all layers improve recognition accuracy in scene text recognition?

Key findings

• FDRN achieves state-of-the-art performance on multiple benchmark datasets for scene text recognition, outperforming existing models.
• The proposed f-RDB reduces computational cost while preserving the local feature fusion and residual learning capabilities of standard residual dense blocks.
• The global dense block (GDB) enables adaptive, holistic learning of global residual features, significantly improving feature representation.
• FDRN demonstrates superior generalization and feature flow across all layers, leading to enhanced recognition accuracy.
• Extensive ablation studies confirm the effectiveness of both f-RDB and GDB components in boosting performance without increasing inference complexity.