Improving The Performance of Handwritten Digits Recognizing Using Convolutional Neural Network with Augmentation Techniques
Annotatsiya
Generally, human beings can see and feel everything around the world with their eyes, and handwritten digits images can be recognized very easily by humans in any format. But the most formatted handwritten digits images only can be easily recognized by the computer. The purpose of this article is to deal with some methods so that the computer understands the irregular handwritten digits. More research articles have been published in this regard. However, most researchers could not obtain higher accuracy results due to the overfitting problem. Using CNN, automatic extraction of distinctive features can more accurately predict the handwritten digits effectively. The goal is to propose a new idea to alter the training data with small transitions to improve the performance of handwritten digit recognition using CNN with augmentation techniques. We used different transitions methods of the image in the augmentation techniques was even possible to easily recognize irregularly handwritten digits and overcome the limited diversity and quantity of images. We were able to get higher accuracy results, up to 99.67%. The MNIST dataset was used for this research. The neural network trained various (28 × 28) pixel size images in the dataset. So, there were 784 neurons as input in our first layer of the neural network. We categorized these images into ten classes in the output layer of our network from 0 to 9. We applied different augmentation techniques by using random transitions such as zoom in, zoom out, image rotation, and horizontal or vertical flip to overcome the limited diversity and quantity of images. And we achieved the handwritten digits recognition accuracy of 96% before the augmentation technique and 99.67% after applying such techniques.