Thesis (M.Sc.) - Cairo University - Faculty of Engineering - Department of Electronics and Communications

This research solves problems of text detection, verification, segmentation, and enhancement in text imagery applications like news and films. Recent approaches are applied in an efficient way. In news videos, locating multiple captions is done using edge detection by grayscale-based and color-based techniques. Stationary as well as moving captions across frames are automatically classified as horizontal or vertical motion using combinatory techniques of recurrent neural network and correlation-based technique.The Convolutional Neural Nets (CNNs) is used to verify the caption as a caption containing text for further processing. In films, several CNNs are implemented to detect frames containing text with high accuracy. Error handling and correction algorithm are applied to resolve classification problems. Multiple frames integration technique is used to extract inserted text in graphics and enhance it. The Correctly Detected Characters (CDC) overall average weighted accuracy for news text recognition using Autoencoder Neural Network (ANN) is 96.07% while the CDC average weighted accuracy for films text translation is 97.79%

Issued also as CD