high performance computing on graphics processing units: hgpu.org

A new approach of fast color character extraction was proposed. Clustering algorithm was adopted in our method to differentiate between objective character regions and background regions on the premise that character regions are nearly monochromatic. However, the key point of this approach was how to select suitable elements’ features based upon the original image information and character’s information; it directly determined extraction results. We brought forward a means for feature selection to resolve clustering’s key problem, and proved the means is effective. Furthermore, in order to overcome the non-ignorable computational overhead of clustering, parallel clustering algorithm was designed and realized by CUDA. Experiments illustrated color character could be extracted satisfactorily even if these characters are surrounded by complex background; meanwhile, parallel improved method that executed by assist of GPU was more efficient than former method that executed by CPU only.