← Projects

Final Year Project · SCSE22-0243

FYP — Self-supervised Model for Image Prediction

NTU SCSE final-year project on self-supervised missing-patch prediction: a convolutional autoencoder with skip connections learns to restore randomly masked regions from surrounding context (CIFAR-10 style 32×32 pipeline).

Self-supervisedAutoencoderImage InpaintingTensorFlowCIFAR-10
Read thesis PDF

Live demo

Select any image. A random 10×10 patch is removed, the FYP-style model fills it, and patch accuracy is reported.

Loading model…

Original

Masked

Reconstructed

Method

  • Self-supervised objective: predict missing pixels from surrounding context.
  • Architecture: encoder–decoder CNN with skip connections — the same structure as the FYP Keras model in final_file_self_supervised.py.
  • Training setup: random 10×10 masks on 32×32 RGB images.
  • Browser demo loads a lightweight retrain of that architecture (original CIFAR checkpoint wasn’t checked in) and reports patch accuracy from SSIM-style similarity + inverse MSE.