The text appearing in videos provides useful information, which can be exploited for developing automatic video indexing and retrieval systems. In this study, we integrated a heuristic and a deep learning-based method using Convolutional Neural Network (CNN) for automatic text extraction from videos. The two independent steps used for text extraction are; candidate text region detection and classification. In first step, rectangular regions were detected that potentially contain text by applying heuristics, which includes morphological processing and geometrical constraints. Then, the obtained candidate text regions were passed through several layers of CNN, that first produced convolutional feature map and then classified the candidate regions into either text or not-text classes. A dataset was prepared by collecting videos from various Turkish channels. 70% of the data was used to train the network while 30% for validation. Experiments showed that our proposed method achieved state-of-the-art performance on our dataset.