Detection of textual data from scene text images is a very thought-provoking issue in computer graphics and visualization. The challenges of text detection come from the nature of images used to detect text from them. The technology proposed is comprised of three main stages: Blurring, transformation to the frequency domain, and de-blurring in the frequency domain as preprocessing procedures; a proposed hybrid model for text detection using k-means clustering and text bounding boxes is used, and (CNN) framework used to classify into textual and non-textual regions. We used the Stroke Width Technique (SWT) when increasing the value of k by analyzing the width of the strokes in the image and thus reducing the potential textual regions resulting from the MSER, which leads to reducing the time and increasing the accuracy of detection, a convolutional neural network (CNN) algorithm is used to perform a classification of all bounding boxes. For the suggested technique's evaluation, results are collected on three popular datasets, including SVT, IIIT5K, ICDAR 2013, and ICDAR 2015. The datasets for SVT, IIIT5K, ICDAR 2013, and ICDAR 2015 each have classification results of 87.41.