Patch loader
Pairs LR/HR files by filename, samples aligned patches, and applies flips/rotations.
PyTorch / Computer Vision / Super-Resolution
Recovering plant leaf detail from low-resolution imagery.
LeafSR packages a Kaggle-style notebook into a reproducible ML system: paired image loading, ESRGAN-lite training, validation metrics, inference scripts, and a visual project frontend.
Validation MAE
16.89
from 18.20 bicubic baseline
Scale
4x
Train pairs
1,642
Test images
495
TTA
8-way
Dataset Explorer
Real paired samples from the competition data.
The frontend commits only a few lightweight sample images. The full dataset, VGG weights, checkpoints, and submissions stay out of git.
Low-res input
Bicubic baseline
LeafSR output
High-res target
Bicubic sample MAE
22.21
LeafSR sample MAE
21.37
Improvement
3.8%
System Design
Notebook logic converted into a reusable training pipeline.
ESRGAN-lite generator
Uses RRDB blocks and pixel shuffle upsampling to predict a residual over bicubic HR.
Hybrid objective
Combines Charbonnier, residual, VGG perceptual, FFT, and low-weight adversarial losses.
Inference
Runs EMA checkpoints with 8-way test-time augmentation for submission generation.
Architecture
Residual super-resolution over a strong classical baseline.
Low-res leaf
Bicubic 4x
RRDB generator predicts residual
LeafSR output
Measured Result
Clear validation improvement over bicubic interpolation.
The trained checkpoint used a Tesla T4 with automatic mixed precision, cosine scheduling, early stopping, EMA-smoothed generator weights, and 8-way test-time augmentation.
Bicubic baseline
18.20 MAE
LeafSR ESRGAN-lite
16.89 MAE
Relative reduction
7.2%