Abstract: Text plays a pivotal role in the images. Text in images accommodate some useful raw facts about the scene and surroundings. Therefore, it is important to detect and recognize them. Some of the limitations to detect texts are intensity, fonts and sometimes even because of complex background. In order to overcome these limitations, first sharpen and blur the image using Gaussian blur and low pass filters, then all the covariant regions are detected. Using all the covariant regions, convolute the image with the Gaussian kernel and continuously down sample the image by half. After convoluting, create a feature map and apply morphological gradient to the resultant image in order to cluster the text regions. After doing this, the non-text regions are eliminated and the text regions are detected. The proposed approach is experimented on ICDAR 2015 dataset and the results show that the proposed approach is robust and it outperforms for various kinds of images.
Keywords: Gaussian kernel, text detection, morphological gradient.