Research On Translation-rotation And Feature Extraction Of Images In Handwritten Digits Recognition

With the development of artificial intelligence and machine learning,people began to use Neural Networks and other frames to make computers imitate the learning mechanism of the human brain and learn from the training data,and then solved complex problems.Thereinto,handwritten digits recognition has attracted more and more attention because of its wide application prospects and good commercial value,such as verification of bank bills,recognition of license plate numbers and check of personal identity.However,in the traditional BP(Back Propagation)Neural Network,we found two problems: one is poor translation-rotation stability of networks.The recognition accuracy rate will drop significantly while translating or rotating test images which is obviously different from the handwritten digits recognition of human brain system.The other one is too many input parameters of networks.The input parameters of BP Neural Network are generally all pixels of each digit image.Because of too many inputs and training parameters,the training time is long and the efficiency is low.In this thesis,for the problem of poor translation-rotation stability,we propose to add two pre-processing layers based on BP Neural Network: for the translation problem,we propose to add a sliding window to center each image to uniform location information of digits;for the rotation problem,we propose to determine the principal direction of each image according to PCA(Principal Component Analysis)and align its axes to the principal direction to uniform direction information of digits.The network improved not only enhances translation-rotation stability significantly,but also increases the recognition accuracy rate because it removes the interference from different digital positions and directions.In this thesis,for the problem of too many input parameters,we propose to extract the feature of each block for images to simplify the network structure and improve the performance of neural networks.In this thesis,the methods proposed of extracting features include: counting the area of the digit contours,calculating the center of gravity for images and calculating image moments.These features extracted replace the pixels into the networks.While selecting the center of gravity or image moments as input,we not only reduce the training parameters of networks,but alsoimprove the accuracy rate of digits recognition.In this thesis,the research on neural networks is not confined to the current main research direction,that is,changing the network structure to solve the problem through the complex connection and learning algorithms.Instead,we hope to remove worthless and interfered information,and extract useful information in order to improve performance of networks by pre-processing the input data based on a simple network model.