Handwritten Character Handle Based On CUDA And Deep Belief Networks

In order to process massive amounts of handwritten character recognition, recognition rate should increase gradually due to the huge data base. This article mainly increased the recognition speed for handwritten character and writer identification content. In the first section, Deep Belief Networks are utilized to recognize handwritten characters. Due to the multi-layer of networks, however, the required speed in practice cannot be reach because of the onerous calculation. In the second section, feature extraction based on microstructural character value is utilized for the recognition of writer identification content. The extraction of the microstructural character value did plenty of calculation, which is way more than common recognition methods, therefore the recognition result is better and it’s limited in small amount of recognition. NVIDIA’s GPU (Graphic Processing Unit), which has the parallel programing ability, is used as the parallel processor. It can process the recognition parallel, and make sure that the neural cells in the networks are paralleled. For the two sections described above, with the GPU multi-picture can be processed simultaneously and the recognition rate for the handwritten characters can be ensured, the calculation speed for the process of multiple feature vector is increased concurrently. In the end, the differences in recognition rate and accuracy between the traditional methods and the ones used in this article are discussed. With no influence on accuracy, the recognition speed is improved substantially.