Character Recognition using Adjustment Convolutional Network with Dropout Layer

Character Recognition is the machine's ability to receive and translate certain character input. It can be handwriting, computer-fonts, and images. In this article, character recognition is done using the Adjustment Convolution Network (ACN). The ACN architecture used consists of 28 layers (input, 5 convolutional, 5 Batch Normalization, 5 Rectified Linear Units, 4 max-pooling, 5 dropouts, Fully Connected Layer, Softmax, Output). Dropout is located in each convolution block. The values used are 0.25, 0.3 and 0.4. Dropout keeps away from the overfitting incident. The network is carried out with the optimization of Adaptive Moment Estimation (Adam) which is a combination of momentum and adaptive sub-gradient variable. Data contain digits 0-9, letters A-Z, and letters a-z that amounted to 171,562. Data is taken from MNIST, notMNIST and Char74K public data. A comparison between training vs. testing data is 70:30. The results showed accuracy of up to 99.04% (MNIST), 94.90% (notMNIST), 98.9% (font-Char74K), 79.4% (Handwritten-Char74K), and 86.80% (Image-Char74K) respectively.