The Torn Page

Seeing what the fire took.

Deep Learning-Based Complete Image Prediction from Partial or Occluded ImagesUndergraduate thesis, defended 2026

27.60 dBPSNR

Defended

Supervised by Dr. Chowdhury Mofizur Rahman, BRAC University.

Back to the journal

The Problem

Given an image with part of it missing or occluded, predict the complete image. The framework does it without a GAN, in a single pass, at roughly 32 ms per 256x256 image.

The Framework

Co-developed with the thesis team, the framework runs in three stages:

  1. Dual-Manifold VQ-VAE
  2. Latent Attention U-Net Mapper
  3. Multi-Scale Detail Injection

What Araf owned

PartWhat it involved
Mask generationThe masks the model learns to fill.
Zero-shot evaluationTesting on four unseen datasets.
BaselinesRetraining five baselines under matched conditions.
AblationsThe study showing a 52.7% cut in mapper validation latent loss.

The Take

Still to come

Real thesis figures and sample input, masked and output images, a link to the paper or report, and possibly an in-browser demo. Until then, the numbers above are the whole account.

Wood engraving: two men rest by a bark lean-to and a smouldering campfire on a lakeshore
Camping Out in the Adirondack Mountains, after Winslow Homer, 1874. A stand-in until the thesis figures are published.

The stack

PyTorch