Text Extraction and Detection from Images using CRNN and Security Algorithm

The rapid digitalization of text information calls for cost-effective and safe optical character recognition (OCR) methods. Herein, the current paper advocates an advanced text detection and text extraction scheme for images using a deep learning architecture incorporating Recurrent Neural Networks (RNN) along with cryptography based on AES. The model improves text recognition and detection accuracy with a hybrid Convolution Neural Network (CNN)-RNN model in conjunction with text localization methods like Efficient and Accurate Scene Text Detector (EAST) and Connectionist Text Proposal Network (CTPN). For maintaining data integrity and security, the extracted text is encrypted with AES, which offers a tamper-proof solution for safe document processing. Experimental evaluation on benchmark datasets (ICDAR 2013, SVT, IIIT 5K) and a self-created dataset of 5,000 images shows superior performance with an attained 5.2% Character Error Rate (CER) and 7.3% Word Error Rate (WER) much higher compared to standard OCR models like Tesseract. The suggested framework also attains a 91.6% F1-score in the ICDAR 2013 dataset and demonstrates strong adaptability for handwritten, printed, and scene text in different environments. Additionally, computational efficiency in the AES encryption mechanism provides little overhead, thereby making the system appropriate for real-time use. This work contributes to OCR technology by tackling both accuracy and security, providing a secure and scalable solution for industries needing automated text recognition, including finance, healthcare, and legal document processing.