Mrs Performance evaluation of Convolution Neural network for handwritten Digit Recognition
T. R. Manjula · SPAST Abstracts · 2021
Image classification problems are very well addressed by computer vision methods. However is not devoid of manual feature extraction. The recent advancements in the field of artificial neural network and particularly convolution neural network are proven to outperform the conventional methods. The CNN, a deep learning technique is capable of addressing a large number of classification and recognition problems. However, there is no one unique model that works for all and exhibits high degree of flexibility in the selection of model parameters such as filter count, kernel size, number of layers, pooling size and optimiser. The performance of CNN is evaluated for handwritten digit recognition of MNIST data base. The kernel size and type of optimiser have greater contribution on accuracy. The single layer CNN model with 32 filter count, kernel size of 9x9, pooling size of 2x2 and adam as an optimiser has achieved an recognition accuracy of 99.13%