A New Image Classification System Using Deep Convolution Neural Network And Modified Amsgrad Optimizer
Arman Mohammed, Ahmed Ak. Tahir · The Journal of The University of Duhok · 2019
A new deep Convolutional Neural Network (CNN) with six convolutional layers and one fully-connected layer is developed and trained by backpropagation using a new optimization algorithm called Fast-AMSgrad which is modified from AMSgrad. The aims are to speed up the training process while achieving acceptable accuracy. The application of the network using both, the Fast-AMSgrad and the AMSgrad algorithms to CIFAR-10 dataset for image classification reveals that the developed CNN performs better when trained with Fast-AMSgrad for both cases, with and without Batch Normalization (BN) layers. The training time is reduced by 50% when Fast-AMSgrad algorithm is used. Also the accuracy and loss values of the training and validation are improved when Fast-AMSgrad is used. The training and validation accuracies provided by Fast-AMSgrad with BN are (91.18% and 86.92%) at epoch number (50) and (94.13% and 86.758%) at epoch number (100), while the corresponding accuracies that are provided by AMSgrad with BN are (82.65% and 81.4%) at epoch (50) and (88.82% and 85.85%) at epoch (100). The overall test accuracy and classification metric measures indicate that the given architecture of CNN and optimization algorithm perform reasonably well