The Dogs-vs-Cats dataset from the Kaggle computer vision competition was used to train a deep neural network model, predicting the probability of the image being a dog or a cat using a pretrained convnet model.
A newly created small dataset consists of 2,000 images of dogs and cats (1000 from each class) is used to train the model. It contains three subsets: a training set with 1,000 samples chosen from each class, a validation set with 500 samples chosen from each class, and a test set with 500 samples.
The pre-built model uses R's keras package to build a convolutional neural network (CNN). The knowledge is discovered using a keras sequential model constructed by using a pretrained vgg16 model as the convolutional base and adding dense layers on top. A mathematical measure of the information content of the model is used to guide the network construction. CNNs are a popular deep neural network learner for image classification because they are highly reusable.
The demo command applies the pre-built model to a test dataset with known values of the labels and presents an evaluation of the performance of the model in terms of loss, accuracy, and confusion matrix. Errors in the predictions are highlighted in the output. The model using feature extraction with a pretrained network results in an accuracy of 90% on the test set from Kaggle. Here, we replace the test set with 50 samples of each class from PEXELS for demonstraction purpose to avoid licensing issue.
The print command displays a textual summary of the model and its model parameters. A textual presentation is not always easy to read and so the display command will draw the model as a network plot.
The score command launches a shiny app to interactively provide new image from the category of dogs and cats to obtain the prediction of belonging to dogs or cats.
You can retrain the model on your own labeled image datasets (training and validation data). Feature extraction with data augmentation is a valuable technique for working with small image datasets, as well as a powerful way to fight overfitting. The model performance could be further improved by fine-tuning some of the layers.
This image classificaition example could be easily adapted to other types of images.
You can go through the full tutorial keras-image-classification-on-small-datasets on Rstudio tensorflow website.