Dense Residual Network: Enhancing global dense feature flow for character recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep Convolutional Neural Networks (CNNs), like Dense Convolutional Network (DenseNet), excel at image representation learning through hierarchical features.
• Existing CNN architectures often struggle to fully exploit local and global feature information across layers.

Purpose of the Study:

• To enhance the local and global feature learning capabilities of DenseNet.
• To improve image representation by fully exploiting hierarchical features from all convolutional layers.

Main Methods:

• Proposed a Dense Residual Network (DRN) for optical character recognition.
• Introduced a refined residual dense block (r-RDB) for local feature fusion and residual learning.
• Developed a global dense block (GDB) using r-RDBs and sum operation for adaptive global feature learning.
• Incorporated a down-sampling block with convolutional layers to reduce feature size and extract deeper features.

Main Results:

• The proposed DRN model demonstrated enhanced performance in optical character recognition tasks.
• DRN effectively captures both local and global hierarchical features.
• The refined blocks (r-RDB and GDB) contribute to improved feature learning and computational efficiency.

Conclusions:

• The Dense Residual Network (DRN) offers an effective approach to enhance feature learning in deep convolutional networks.
• DRN provides superior results compared to existing deep models for optical character recognition.
• The proposed architecture successfully addresses limitations in exploiting hierarchical features for improved image representation.