Intervention of light convolutional neural network in document survey form processing.

Public survey is popular in different domain to have a feedback from service users in the aim to improve the quality of service provided or to figure out what they think. In general, the data processing takes time due of manual intervention and it may distort the result because an error of processing may occur, so the aim of this research is to propose a new finding how to process rapidly a paper survey form to have the statistic result on time. The paper template design must have at least one question with "Yes" and "No" check box or radio button as response field. People have a liberal choice to reply: normal, no response or both. In each box in form can contain a different type of handwriting and the way to recognize them is within mathematics tools combined with neural networks. First of all, the physical paper which contents the information must be converted to image by digitization operation. Any type format of image is accepted then it will be handled by preprocessing module before coming to the Lite Convolutional Neural Network. Once the neural networks finish the treatment, a postprocessing module computes the information and insert into database. The singularity of this technique is the intervention of Lite Convolutional Neural Network as recognition tool which has more advantage compared with standard deep learning in terms of computation time, accuracy/precision and number of parameters. The validation of this approach is done with RVL-CDIP dataset, real survey of mobile money and satellite TV service, the accuracy reaches 96%, 97% and 98% respectively. [ABSTRACT FROM AUTHOR]