Abstarct: This thesis aims at finding the text regions in Farsi natural scene text images. Text Localization is defined as locating regions in images that are considered as text by human beings. This problem is challenging and open because text can vary in size, font, and color, and the background can be non-uniform. As the Farsi text is intertwined and very similar to foliage, it is even more challenging to locate. Four distinguished methods for localization of Farsi and Latin text are introduced in this thesis. Each of these methods addresses the weak points of the previous ones and improve upon them. The first method, Edge Color Signature, introduces colored edges for locating candidate text regions. First, the Mean-Shift clustering algorithm is employed by this method for assigning colors to edge pixels. Consequently, these colored edges are utilized to produce candidate text regions. Finally, a cascade classifier is used to classify the candidate regions. This classifier employs novel features baxsed on edge pixels and dictionary learning methods to detect text regions. The second method, Edge Color Transform, improves the edge extraction method of Edge Color Signature. Also, it introduces a novel, fast and accurate operator for producing colored edges. The third method, Deep Color Transform, improves the second method to locate text with non-uniform backgrounds. Also, a convolutional neural network with novel architecture is introduced to classify the candidate regions. Finally, the fourth method, called mexta-Heuristic Net, uses the convolutional network of the third method; The network's output is employed in a mexta-heuristic search method to locate text regions. In this search method, search windows are distributed over the input image and gradually converge toward the text regions. The proposed methods are examined and compared with each other and several state-of-the-art methods on two datasets: Farset for Farsi text and ICDAR2013 for English text. The results show that the Deep Color Transform with the f-measure value of 86.64 for English and 91.58 for Farsi images outperforms all of the other methods.