The Research Of Handwriting Score Recognition In Test Paper Based On Convolutional Neural Network

In the modern teaching process,the automatic recognition and input of students' test scores are important for the standardization of school teaching management.The automatic identification of test scores can effectively reduce the workload of teachers,and thus have more energy to devote to teaching and research work.Handwritten digit recognition is the key to recognize and input test scores automatically.Therefore,aiming at the recognition of handwritten score,this paper analyses the existing algorithms,and proposes two kinds of handwritten digital recognition algorithms based on convolution neural network(CNN).The main work and novelty are summarized as follows.Firstly,the paper uses the image segmentation technology based on HSV color space to segment the handwritten scores characters,and then combines the projection method to extract the handwritten score area.At the same time,the projection method and the drip algorithm are used to divide the scores characters of each question,including the segmentation of joined-up and tilted digital characters.In this paper,the problem of handwritten test scores recognition mainly includes two parts: image preprocessing and handwritten digit recognition.The extraction of handwritten scores area and digital character segmentation are important parts of image preprocessing.Secondly,in order to improve the recognition performance of CNN model,a handwritten digit recognition algorithm based on improving CNN is proposed.At first,the paper,combining the advantages of the two activation functions of Relu and Softplus,comes up with a piecewise activation function,which has the characteristics of sparse and smooth.Then,the mixed sampling method,taking the eigenvalues extracted both from mean pooling method and the maximum pooling method into account,is proposed during the downsampling operation,so that the model will have smaller error and higher stability.Finally,the method of increasing the momentum items and the adjusting strategy of learning rate self-adapting are used to improve the network training speed in the network training stage.Furthermore,a Dropout is added behind the CNN full connection layer to prevent the network from overfitting.Experimental results show that the proposed algorithm has better recognition performance than the traditional CNN model and other methods,and the simulation results on the MNIST data set fully verify the effectiveness of the proposed algorithm.Thirdly,in order to fully extract the image information,a handwritten digit recognition algorithm based on the feature fusion and SVM is proposed.First,the CNN features and Gabor features of character images are extracted by using the improved CNN model and the Gabor filter with the curvature coefficient.Then,fuse the features extracted by CNN and Gabor to get more effective new features.Finally,the author input the fusion feature into SVM classifier for handwritten digit recognition.The experimental results show that the proposed algorithm can effectively improve the recognition effect of handwritten digits.