Using Tensorflow and keras for building CNN model with CIFAR-10 dataset for image classification
Built different CNN models using different techniques and layers
Achieved an accuracy of more than 85% between train and test sets
Final model gave an accuracy of 85%
Used 3 convolution layers with MaxPooling, 3 Flattening Dense layers, ‘Adam’ optimizer, ‘relu’ and ‘LeakyReLU ‘ as activation functions
BatchNormalization and Dropout were used in different layers
Model was run with batch size =10 and epochs = 100
Model stabilized by 20 epochs (validation set) with reducing marginal improvement
Model can be improved by using more convolution layers, different optimizers, activation functions, epochs and transfer learning. But that will require more compute.
You can try using different image datasets along with transfer learning to improve the model. Rest is all about hypertuning NN layers and their parameters.
Code in github
https://github.com/datawisdomx/ComputerVision_CNN_CIFAR10_ImageClassification_Tensorflow
