Transfer learning based Optical Character Recognition using Natural Images
Publication Date : 23/12/2019
Text detection and recognition has been a well-studied problem in the past. However, when it comes to detection and identification of text in natural images and text from natural scenes, it becomes a much more challenging problem because of the distortion in geometry, variance in the illumination. Recently, deep learning techniques achieved state-of-the-art performance in object detection and recognition. But deep learning required large data set and high computation power for training model from the scratch. Therefore, in this research work, a deep learning technique is used, which is based on the knowledge transfer of pre-trained Convolutional neural networks (CNNs) to recognize the text in the natural images. In this work we have uses pre-trained VGG19 model. To fine-tuned the VGG19 model top layer is removed and new layer have been include in the network then it is trained (fine-tuned) on the optical character images. A CHAR74K dataset, which is a benchmark for natural images, is used for the evaluation of the proposed method. The proposed model has achieved an accuracy of 87.56 % and F1- score of 88%.
No. of Downloads :