Mirror Text Classification from Image Using Machine Learning Techniques

Text classification in natural scene photos is a crucial precondition of numerous content-based image analysis applications, even though the majority of recent research works primarily focus on horizontal or nearly horizontal picture text. Some of the work on Multi-orientation Scene Text Detection that has been done. There hasn’t been any research done on the detection of mirror text. In the article, an efficient method for identification of mirror text from image using machine learning algorithms is proposed. The RGB input images are first converted to grayscale, then edge detection using the Sobel edge detector, histogram equalization to improve the quality of the grayscale image, then dynamic thresholding to turn the image into a binary image. The largest possible rectangular contour area is used by the machine for localizing the text. The text words are arranged together by the system into various clusters. Sub-clustering creates distinct sub-clusters for each character and is applied to particular clusters. The system extracts the vector contour (VC) for each class. To assign every sub-cluster to the individual class that was previously determined, the system uses the inter co-relation function (ICF). It accomplishes this by determining the highest similarity between test and previously trained VCs. If one or more mirror character found in the text it classify that text as mirror text. The system is trained on 62 mirror character classes, with 100 samples created via the augmentation procedure in each class. Another set of 62 classes with a total of 62 \(\,\times \,\) 100 = 6200 photos is used to evaluate the system, yielding a recognition accuracy of 84%.