Handwritten Numeral Recognition And Test-Paper Management Based On Neural Network And GPU

The digital management of test paper has a dramatic effect on improving educational quality. The information about students'learning on knowledge can be gained effectively by using a manager system of test paper. Therefore, a digital manager system of test paper , which has a prospect in practice, was designed here.Two key points in the management of test paper are to realize digital recognition of test paper and structural analysis. Among the process, Automatic recognition of student ID and scores is a critical step. To realize the application in practical work of this technic, BP neural network was adapted for the training and recognizing algorithm in our research. At last, a higher speed of BP neural network training and recognizing process was achieved based on GPU, so that the time for training neural net was shorter than before. An overall frame of test paper manager system was designed and several core functions were realized, for example, the numerical recognition of student ID. Four main parts of our work were as follows:Firstly, the pretreatment was realized by image grayness,Image Binary and Image denosing, so that binary Images which were easy to recognize were achieved. The area of blank was located by projection method to fix the position of area with numbers. At last, the recognited-numbers were gained by dividing and normalizing.Secondly, BP neural net was proposed for numerical recognition after considering the characteristics of digital management of test paper. This method was applied in the test paper manger system successfully.Thirdly, the efficiency of hand-writing numerical recognition by using neural net was closely relevant to the size of training sample. Time for training would take longer time as the size of sample became bigger. Based the situation, the accelerating process of training for neural net by using GPU was analyzed here.Finally, a manger system of test paper was realized here by using C++, in which BP neural net was applied based on GPU. The system combined hand-writing numerical recognition and test paper management together. Student ID and scores could be stored automatically. Calculation based on GPU was firstly applied in test paper manger system, so that the speed of recognition algorithm was more efficient. The training speed of 3-TIERED neural net after accelerating on GPU was three to five times than before. This paper offered a new solution to practical application of neural net, in which the efficiency was a key problem in sample training with large size.