content‑aware image retargeting
content‑aware retargeting via dynamic programming: upload, set size, preview seams, download.
general info
- tech: Python 3.11, Streamlit, OpenCV, NumPy, Docker
- algorithm: Sobel (grayscale) energy → DP cumulative costs (vertical & horizontal) → seam backtracking → seam removal (vectorized)
- UI: image uploader, width/height sliders, centered non‑stretch rendering, 3‑tile grids for original and resized seam previews, persistent download
- shape: small library in
src/seam_carving/ + modular UI in src/ui/ + Streamlit entry in src/main.py
why it matters
uniform scaling squashes important content. seam carving removes low‑energy paths to better preserve salient regions. this project pairs a readable implementation with an interactive UI to make the idea tangible.
architecture
algorithm
- energy: grayscale Sobel gradients, summed magnitude
- DP (vectorized): for each row/column, pick predecessor with minimal cumulative cost; store step deltas in {-1, 0, +1}
- backtracking: trace minimal seam from the last row/column using stored step deltas
- seam removal: boolean‑mask + reshape (vectorized) for vertical and horizontal seams
- scheduling: interleave vertical/horizontal removals (Bresenham‑like) to reduce distortion
data flow
- upload → decode (BGR→RGB) → show next seams on original
- sliders select target size (must be ≤ original)
- carve → store result in session → show result + next seams → download PNG (persists on rerun)
notes
- currently supports shrinking only (no expansion)
- performance scales with pixels × seams; consider pre‑resizing very large images
- Sobel energy is fast and effective; swapping in Scharr/forward energy is a natural extension
- next seams are previews; they change after each removal by design