An innovative network based on double receptive field and Recursive Bi-directional Long Short-Term Memory

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Sequence recognition in natural scene images is crucial for computer vision.
• Convolutional Recurrent Neural Networks (CRNN) are popular for end-to-end character recognition.
• Existing CRNN models struggle with wide characters and recognizing long, dense, small characters.

Purpose of the Study:

• To address the limitations of CRNN for recognizing challenging character sequences.
• To propose an improved CRNN network, CRNN-RES, enhancing feature extraction and parameter efficiency.
• To improve the accuracy of character sequence recognition in natural scene images.

Main Methods:

• Developed CRNN-RES, an improved CRNN architecture.
• Integrated a dual pooling core within the CNN for enhanced feature extraction.
• Modified the RNN layer to a shared parameter BiLSTM network with recursive residuals.
• Introduced the CRFC layer for flexible input sequence length configuration.

Main Results:

• CRNN-RES achieved higher accuracy across multiple public datasets (ICDAR 2003, 2013, IIIT 5k, SVT) compared to the original CRNN.
• CRNN-RES demonstrated improved performance, with accuracy gains of 1.40% to 5.43%.
• The CRNN-RES network reduced parameter count by 133,752 compared to CRNN (8,197,549 total parameters).

Conclusions:

• CRNN-RES effectively overcomes CRNN's limitations in recognizing long, dense, small characters.
• The proposed architecture offers a more parameter-efficient and accurate solution for natural scene character sequence recognition.
• CRNN-RES shows significant potential for real-world applications requiring robust text recognition.