Image-Classification-Using-CIFAR10

V1 - Using Random Forests

• The first version of image classification on this dataset has been done using Random Forests Classifier.The classifier gave an accuracy of 46.72%(Scope of improvement still there)

V2 - Using Convolutional Neural Networks

• This version of image classification on CIFAR-10 uses 3-layers of CNN along with 1-Dense layer.
• The code has been written using the Keras deep learning library with a Tensorflow backend.
• An accuracy of 84% is achieved on Train dataset and of 79.9% on the Test dataset.
• A significant gain is seen from the earlier version of image classification which used the Random Forest classifier.

V3 - Using Convolutional Neural Networks along with data augmentation

• This version uses the Image Generator function inside Keras for data augmentation.
• A lower accuracy is achieved on the train dataset(75%) and a higher accuracy(79%) on test dataset.
• Test dataset is different from the train dataset and this helps in achieving this higher dataset.
• A different model architecture is used to get better results(6 CNN Layers with 1-Dense Layer).

V4 - Using Transfer Learning

• The VGG16 model is used along with image_net weights.
• On training the model using just the CNN-layers from VGG16 along with a softmax output layer, an accuracy of 67% was achieved on the test set.
• Using a 1024 units dense layer after the VGG16 input, gave an accuracy of 71%.
• The model can give much higher accuracy on performing fine tuning.