Week 7# Identification of Artists and Movements from Paintings with Machine Learning

This week we are publishing our last blog post and we will present the results of AlexNet.

AlexNet:

AlexNet is a eight-layer convolutional neural network that can achieve high accuracies on even the most difficult datasets. First five layers are convolutional layers and followed by max pooling layers. Last three are fully connected layers. See below figure for the AlexNet’s architecture:

Figure1: AlexNet’s Architecture

While testing the AlexNet we tried many combinations. Batch size was between 1–1024, as an activation function we tried RELU, ELU, SELU, tanh and epoch of 300. We get the best results with Batch size = 32, epoch = 50 and SELU as an activation function. See the confusion matrix and accuracy table below for the results:

Figure2: Confusion Matrix of Artist Identification (Model: AlexNet, Epoch = 50, Activation Function = SELU, Batch Size = 64)

If we give an example where we misclassified artists, most noticable one is, missclassification between “Claude Monet” and “Pierre Aguste Renoir”. Two artists gave works on Impressionism Art Movement. See below example for misclassification:

Figure3: Left(Pierre Auguste Renoir) — Right(Claude Monet)

While classifying the movements, we also used the same hyperparameters with artist identification. See the confusion matrix and accuracy table below for the results:

Figure4: Confusion Matrix of Movement Identification (Model: AlexNet, Epoch = 50, Activation Function = SELU, Batch Size = 64)

If we give an example where we misclassified movements most noticable one is, missclassification between “Symbolism” and “Impressionism” just like in LeNet-5 model. See below example for misclassification:

Figure5: Left(Impressionism) — Right(Symbolism)

Overall we got 80.44% accuracy for artist identification and 83.10% accuracy for movement identification in our AlexNet model. It is better than LeNet-5(was 70.22% accuracy for artist and 76.93% accuracy for movement identification).

Thank you!