| As an important part in the stage of primary education,primary arithmetic plays an important role in cultivating students’ basic mathematical literacy and enhancing their competitiveness in this stage.However,the training of primary arithmetic requires students carry on a lot of practice and examination,which will bring a lot of manpower cost.Therefore,this paper proposes to use artificial intelligence technology to recognize those papers,in order to reduce the workload of teachers.In this paper,the whole recognition process is divided into three parts,that is,the scan of the examination paper,the location of the primary arithmetic questions and the recognition of the primary arithmetic questions.Considering that the scanned images may still be inclined and there may be too much irrelevant information in it too that will subsequent the positioning and recognition operations.At the same time,the sample sets collected from cooperative schools will also face the problems of insufficient quantity and uneven distribution.Secondly,due to problem of the handwritten part of the elementary arithmetic questions including the large difference in character shape and a large number of non-standard writing,that will be a huge challenge for the positioning module and recognition module in the automatic marking system.If it is only a rigid copy of the current common model for positioning and recognition operation,the results will be less than satisfactory.In view of the above problems,the main research work of this paper is as follows:1.Considering the scanned image may exist a series of problems that affect the positioning.This paper has proposed some pretreatment methods such as binarization,inclined correction and segmentation.At the same time,this paper has proposed a training sample generation method aiming at the problems existing in the training sample set of recognition module to ensure the recognition effect of the final model.2.To avoid the non-specific writing of the handwritten character in primary arithmetic questions will seriously affected the position effect of positioning module.This paper has selected the Faster R-CNN network based on the FPN structure as the basic network in the positioning module and has selected the Res Net network for feature extraction,so as to fully extract the different features of the image.At the same time,this paper has replaced some parts of the convolution layer in the Res Net network with deformable convolution in order to better extract features for primary arithmetic with different shapes and avoid the interface of irrelevant noise information.3.Considering the problems such as the great difference between printed characters and handwritten characters and the non-specific writing of handwritten characters will be seriously affected the recognition effect of the recognition module.Therefore,this paper selects Densenet network for feature extraction and improves it into a multi-dimensional feature output structure,so that to extract all kinds of features needed for recognition from primary arithmetic question images as much as possible.At the same time,this paper has selected the encoder-decoder model based on the joint CTC-Attention model and improved it into a structure with multiple input processing,so that to make full use of the multi features extracted by the improved Densenet. |
