GPU-Based Parallel Programming for Digital Document Image Classification

The use of document images is currently increasing. It creates another need to classify document images into their type automatically. To recognize the document images and classify the type, several feature extraction methods have been used in this research. Binary Morphological erosion with 9 intersection types is one of the simple and effective feature extraction method. Due to time processing, this method is not applicable. Therefore, the purpose of this research is focusing on the use of multithreading in multicore processor with POSIX Threads (Phtreads) and multithreading GPUbased with NVIDIA CUDA in order to speed up the processing time. The results show that with local processing, the classification accuracy can be improved. Furthermore, Pthreads implemation in dual core processor can increase the processing speed around 2 scale factor. In the other hand, CUDA implementation can increase the processing speed around 100 scale factor compare with single thread program and 60 scale factor compare with multithreading in dual-core processor.