A Sequential Handwriting Recognition Model Based on a Dynamically Configurable CRNN.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Handwriting recognition is crucial for processing handwritten inputs like text and digits from images.
• Real-world applications often involve sequential text in diverse languages, necessitating dynamic recognition systems.
• Existing methods may lack adaptability for complex, sequential handwritten data.

Purpose of the Study:

• To propose a novel Dynamically Configurable Convolutional Recurrent Neural Network (DC-CRNN) for handwriting recognition.
• To optimize the structure and hyperparameters of Convolutional Recurrent Neural Networks (CRNNs) using metaheuristic algorithms.
• To enhance the performance of handwriting recognition systems for sequential text in multiple languages.

Main Methods:

• Developed a DC-CRNN inspired by neuroevolutionary techniques.
• Utilized the Salp Swarm Optimization Algorithm (SSA) for optimizing CRNN architecture and hyperparameters.
• Investigated two encoding techniques for translating optimization outputs to CRNN recognizers.
• Introduced a hybridized SSA with Late Acceptance Hill-Climbing (LAHC) to refine the optimization process.

Main Results:

• The hybridized SSA with LAHC demonstrated significant improvements in the search process.
• Experimental results on IAM and IFN/ENIT datasets (Arabic and English) validated the proposed approach.
• The DC-CRNN model achieved superior performance compared to handcrafted CRNN methods.

Conclusions:

• The proposed DC-CRNN, optimized via a hybrid SSA-LAHC algorithm, offers enhanced handwriting recognition capabilities.
• The method effectively models sequential handwritten text across different languages.
• This approach represents a significant advancement over traditional, non-configurable CRNN models.