Abstract: This paper presents a detailed framework for an Optical Character Recognition (OCR) system employing Convolutional Neural Networks (CNNs) for recognizing optical characters. The CNN architecture demonstrates remarkable proficiency in learning various styles within input images, including handwriting and printed text. CNNs, as a subset of Deep Neural Networks, excel in recognizing and classifying specific features from images, making them widely applicable in visual image analysis tasks such as image classification, medical image analysis, and language processing.
The paper outlines the essential modules and algorithms utilized in the implementation process. These modules include image processing, segmentation, feature extraction, and training/recognition. In the image processing module, steps such as grey scale conversion and image binarization are employed to prepare the input image for segmentation. Segmentation is achieved through line segmentation, word segmentation, and character segmentation, facilitating the extraction of individual characters from the document image. Feature extraction involves resizing character images and storing extracted features for further processing. Finally, the training and recognition module utilizes the Kohonen algorithm, based on Self-Organizing Maps, for training and recognizing characters.
By presenting this comprehensive framework, the paper aims to contribute to the advancement of OCR systems, particularly in the context of document digitization and text recognition tasks. The proposed approach offers a robust methodology for accurately extracting and recognizing characters from various types of documents, thus facilitating automation and efficiency in document processing tasks.
Keywords: Optical Character Recognition, Convolutional Neural Networks, Image Processing, Segmentation, Feature Extraction, Kohonen Algorithm.
|
DOI:
10.17148/IJARCCE.2024.134141
