SVM vs RF vs NeuralNets on fine grain Image Recognition

Dataset: https://www.kaggle.com/c/dog-breed-identification/data

Data and Preliminary analysis:

• Image Transformation:
  • Size normalization
  • Color datasets: colored, black & white, grey scale
• Data Augmentation: horizontal flip, random sharpen, lightness, contrast, affination, RGB modification...
• Histogram of Oriented Gradients (HOG): feature extractor to be used in SVM and MLP
• Principal Component Analysis: dimensionality reduction (combined with HOG)

Model Building and Validation:

• HOG + Support Vector Machine (with and without PCA) and stacked Models
• HOG + Random Forests (with and without PCA) and stacked Models
• HOG + Multilayer Perceptron and Stacked Models (minimum loss and max accuracy)
• Convolutional Neural Network and Stacked Models (minimum loss and max accuracy)

Not using pretrained Neural Networks

---

Files:

• SVMRFbreeds: MAIN FILE. SVM and RF algorithms. Python 2.7
• CNNbreeds: Convolutional Neural Net. Python 3.5
• MLPbreeds: MultiLayer Perceptron. Python 3.5
• breedsHelpFunctions2: help functions. Compatible with Python 2.7 and 3.5