Links

• https://derma.jmir.org/2022/3/e39143/

Pipeline

Pre Processing

• remove ruler lines from images
• remove hair from images
  • https://pdf.sciencedirectassets.com/280281/1-s2.0-S1110866525X00049/1-s2.0-S1110866525002373/main.pdf
    • use a VAE to remove the hair?
      • https://pmc.ncbi.nlm.nih.gov/articles/PMC9907627/pdf/SRT-28-445.pdf
      • L_(SSIM) + L_2 loss produces the best results
      • Encoder Architecture
        • Layer 1
          • Kernel Size = 3
          • Activation = ReLU
          • Stride = 2
          • Num Filters = 64 (number of kernels)
        • Layer 2
          • Kernel Size = 3
          • Activation = ReLU
          • Stride = 2
          • Num Filters = 32 (number of kernels)
        • Flatten
        • Compute mean and variance for the distribution, latent dimension = 50
        • Sample from the distribution
        • unflatten
      • Decoder Architecture
        • Layer 1
    • DullRazor Algorithm? (Hair Density Score)
      • Target size (images resized to 256x256)
      • Kernel = 7x7
      • Split at score=0.015 yields 8423 hairy samples, 5596 clean
        • take the bottom 1000 clean images
    • CycleGANs? --> some kind of style transfer
• random crop and rotation
• style transfer to get dark skin
• Challenges
  • hair comes in different color, shapes, textures and density, classical CV can only do so much