Frequency transformation for image recognition and inpainting

This is repo from final project for course Linear Algebra. You can read final report here. For checking final score and plots you can use this colab. Final video presention is available here

Installation

git clone https://github.com/igor185/frequency-for-receptive-field
cd frequency-for-receptive-field
conda env create -f env.yml
conda activate fft

Running

Train and evaluate cifar-10

bash runners/cifar-conv-runner.sh
bash runners/cifar-fourier-runner.sh
bash runners/cifar-wavelet-runner.sh

Results

• Classical classifiers:
  • SVM on mnist 91% accuracy, SVM on fft from mnist 84%
• Deep learning classifiers:
  • Network trained on raw pixel performs better then network train on frequencies from pixels
  • Cifar-10: Network(one resnet block) with fft on deep features performs 3% better then network without fft
• Inpainting results(obtained after running inference from this repo you should follow steps in that repo to obtain same result:
  Input image: Inpainting without fft and dilated conv: Inpainting with dilated conv: Inpainting with fft: