Arbitrarily Oriented Scene Text Detection using SMSER and Connected component analysis

Parveen Sharma, Er. Anupam Garg


In this work, rotation invariant approach has been explored and an effective rotation invariant text detection system has been proposed. In this discrete wavelet transform has been used to get the multi-level feature extraction of the text region as vertical, horizontal and diagonal coefficients provide variation in edge pixels of the text scene image. Further this, detailed and approximation coefficients has been processed separately by using different filtering and segmentation methods i.e. Laplacian filtering and fuzzy clustering etc. to get the text region from the scene image. As candidate pixel generation has been done by these methods which separate the text region from the background, it would be easier to segment the characters by using sMSER region based segmentation method. Further connected component analysis has been carried out in which size, height and intensity of the characters has been used to get the final text. Experimental results shows the effectiveness of the system as variety of scene text images having multi-oriented text in them has been considered for the evaluation of the proposed algorithm

Full Text:



  • There are currently no refbacks.