Research On Handwritten English Letter Recognition Algorithm Based On Deep Learning

Deep learning technology is an important sub-discipline in the field of machine learning.Its predecessor is the neural network.In recent years,due to technological breakthroughs,deep learning has yielded good results in many areas of artificial intelligence such as computer vision,natural language processing,machine translation,unmanned technology and robotics.Image recognition is also a very important research area in deep learning technology.Like many other areas of deep learning,it has encountered two challenges of over-fitting and computational efficiency.This thesis focuses on the computational efficiency of deep learning techniques and convolutional neural networks in the field of image recognition.Although for the vast majority of image recognition tasks,the accuracies can be enhanced by several methods with enough data available.Such as increasing the model?s complexity,using more computing resources,or more powerful specialized computing devices.However,in many real scenarios,for example,in the application of mechanical engineering,we can apply this deep learning algorithm based on neural network to the image recognition algorithm of industrial robots and use this algorithm to sort goods or parts.The available computing resources in industrial robots are limited.Which are limiting the complexity of the models that can be used in the application.We need to improve the computational efficiency,reduce the number of parameters,so that it is possible to apply it to more actual scenes.In order to reduce the computational time and improve the computational efficiency,we propose a variant of the Inception structure.Which is based on the multi-branch Inception structure.We use it to construct a deep convolutional neural network.It achieves better classification results and higher computations effectiveness.The main contents of this thesis are as follows:First of all,by comparing the experiments? resluts,we find that the multi-branch structure represented by Inception can achieve better results than the linear structure represented by VGGNet.Second,we split the convolution kernel in the Inception prototype structure,delete some of the split convolution kernels,and then rearrange the remaining convolution kernels into different branches.We propose a variant of the Inception structure.When stacking Inception variant structures and building a complete convolutional neural network,we add some additional layers and adjust some hyperparameters as appropriate.In the comparative experiment with the prototype,we get the following results.First,in the same software and hardware environment,and the input is also the same,for the handwritten English alphabet data set,the variant structure can obtain more than 92% of the training accuracy and 87% of the verification accuracy,slightly better than the experimental accuracy of the prototype.Second,because the Inception variant structure contains fewer parameters than prototypes,it is more computationally efficient,with an average of 14.898% less training time per prototype than a prototype.