Turkish Text Detection System from Videos Using Machine Learning and Deep Learning Techniques

With the advancement in smart devices and high-speed internet, a continual increase in videos demands an efficient and automatic video indexing and retrieval system. To accomplish it, content-based video indexing is an optimal solution by detecting text in videos. In this study, we proposed a text detection system based on machine learning approaches. We compared conventional machine learning approaches with deep learning method. For deep learning, we implemented Convolutional Neural Network (CNN), while Logistic Regression (LR) and Support Vector Machine (SVM) are employed as conventional machine learning techniques to predict the outcome as text or non-text data. We evaluated the proposed systems on our own dataset obtained from various Turkish videos. LR obtained an overall accuracy of 95.0%, whereas SVM achieved 98.7% while CNN secured 99.8% accuracy. The experimental results show that CNN (deep learning approach) was more effective for our Turkish text dataset as compared to LR and SVM.