CIFAR10-HOG-SVM

Classifier for CIFAR-10. Grayscaling, HOG, PCA, and RBF SVM. 62% test accuracy. This classifier does NOT use any neural network or convolutional filters/layers/kernels.

Running the Code

Google Colab

Add the following code as the first cell:

!pip install scikit-learn-intelex

Then, run the rest of the cells.

Local

Install the following packages:

pip install scikit-learn sklearn-intelex scikit-image numpy matplotlib opencv-python tqdm joblib

Then, run the code.

Dimensionality Reduction

1. Input 32x32x3 => 3072
2. Grayscale 32x32x1 => 1024
3. HOG pixels_per_cell=(8, 8), cells_per_block=(2, 2) => 324
4. PCA 80% explained variance => 66

Configuration and Hyperparameters

• Input -> Grayscaling -> HOG -> PCA -> Stardardization + normalization -> SVM
• Support Vector Machine (SVM) with the radial basis function (KBF) kernel
• C = 10
• Scikit-learn modules are used

Time Complexity

• This training time of this non-linear SVM grows faster than linear time against the number of samples
• Training takes 3 minutes on Google Colab Free
• Scoring against the test dataset takes 1 minuite

Accuracy

• 62% against the test dataset
• Confusion matrix also available
  • Good separation between animals and machineries
  • Rather confused between different types of animals (cats vs deer), and different types of machineries (airplanes vs ships)
  • Confused between cats and dogs