Research And Implementation Of Handwritten Objective Question Letter Recognition Based On Convolutional Neural Network And Transfer Learning

In the process of school examination and homework correction,the technology of using the cursor reader to identify the objective questions on the answer sheet is very mature,However,the style of the answer sheet is changeable,and the high quality requirements for the paper of the test lead to an increase in the cost of printing and test implementation.It is not suitable for widespread use in daily high-frequency teaching tests and homework correction.In order to reduce the cost and simplify the correction process,some high-frequency small-scale examinations and daily homework correction gradually do not use the method of filling in the answer sheet,students only need to write the letters of the objective questions directly on the test paper to answer.in order to realize the automatic recognition of the objective questions by the machine and improve recognition efficiency,this paper carried out a recognition research based on convolutional neural network to recognize the letters of handwritten objective questions in the test paper.The main work is as follows:1、This paper collects pictures of handwritten objective letter and enhances them by cropping 、 scaling 、 and adding noise,and using the one-hot method to label,construct and classify real scene datasets,including three types of neat writing,irregular writing,and interference with images,data enhancement is performed on EMNIST handwritten English letter images by rotating,translating,and flipping,and is used as a pre training dataset for transfer learning.2 、 Using multi label classification to identify,The improved Res Net50 is compared with classical models such as original Res Net34 、 VGG16 and original Res Net50 to verify the recognition effect of the classic model in the handwritten objective question letter scenario,the experiment shows that the improved Res Net50 model performs best in the experimental comparison.3、In order to further improve the effect,considering the unbalanced number of pictures in different scenarios,the large amount of interference information in pictures,and the large diversity of letter writing,this paper proposes the following innovation: First,an optimization-based convolutional neural network model is proposed,and the Inception Hand-STN model is constructed using the Inception structure and the STN spatial attention mechanism to recognize handwritten objective letters.Secondly,in the context of this paper,more attention should be paid to the texture and line features of letters,it is proposed to introduce transfer learning theory in the model training process,and use the enhanced EMNIST handwritten English alphabet dataset as a pre-training dataset,and use various transfer strategies to verify the effect of transfer learning,thirdly,a migration strategy of freezing shallow network parameters and fine-tuning the remaining convolutional layer parameters in the application scenario of this article is proposed.The experimental results show that the maximum recognition m AP value of this strategy under the Inception Hand-STN model reaches 98.9%,which is 4.3% higher than the m AP value of 94.6% without using transfer learning.4、Developed a handwritten objective question letter recognition system,which enables users to make cards 、 recognize and count scores on test papers with handwritten objective questions.Use Py Qt to build a visual operation interface,use Python to complete the information transfer of the main part,and use Opencv image library to develop the image processing part.