Optical Character Recognition Using Optimized Convolutional Networks*
Anum Nawaz, Muhammad Irfan, Tomi Westerlund · 2023
Optical Character Recognition (OCR) has been a prominent area of research in pattern recognition for several decades, owing to its broad application potential in smart living. To improve offline OCR on mobile devices with limited computing resource, we have optimized Convolutional Neural Networks (CNNs) to efficiently detect text using minimal resources. To achieve this, we employed two distinct pretrained CNN models, namely AlexNet and Inception-V3, for feature extraction. Leveraging these models' unique characteristics and capabilities to extract diverse features, we aimed to enhance the classifier's accuracy. This, in turn, facilitates the development of an efficient edge-device application for faster and higher-quality OCR. Experimental results demonstrate that our proposed optimized algorithm outperforms existing CNN-based methods in the field of OCR, particularly in the categorization and detection of handwritten digits and character recognition. The conducted research yielded impressive accuracy results, with up to 97% accuracy on the MNIST dataset and 95.5% accuracy on the NIST dataset.