Geometry comes before blending
Filtering can soften a seam, but it cannot recover consistent geometry from misaligned structures.
Computer
visionPersonal image-processing experiments
Exploring how image geometry, clustering and regression transform pixels into useful results.
- MY ROLE
- Algorithm implementation & experimentation
- CONTEXT
- Personal learning project
- TOOLS
- Python · NumPy · OpenCV
- FOCUS
- Stitching, segmentation & restoration
01 / THE PROJECT
Explore pixels.
Build understanding.
Develop image-processing algorithms and investigate how their assumptions affect the result.
- I undertook this personal learning project to understand computer vision through implementation and visual experimentation.
- The work explored panorama stitching, image segmentation and reconstruction from corrupted images.
- Python notebooks captured intermediate results, parameter comparisons and observations.
3
Featured experiment areas
2
Own photographs used in the panorama
5
Colour and position features
used in superpixel clustering
These figures describe the featured implementation, rather than measured accuracy.
02 / EXPERIMENT WORKFLOW
See the pipeline.
Control each step.
- Each experiment connected input preparation, mathematical operations and visual inspection.
- Intermediate plots made coordinate, intensity and parameter choices easier to investigate.
- Prepare: Load images and establish colour order, intensity scale and pixel coordinates.
- Implement: Write interpolation, clustering and regression calculations in Python.
- Experiment: Vary seam filters, cluster counts, spatial weights and corruption patterns.
- Inspect: Compare intermediate outputs and final results with the original images.
PERSONAL LEARNING GOALS
Questions that guided the work
- How can overlapping photographs be mapped into a common image plane?
- How do interpolation and filtering affect the appearance of a panorama seam?
- How does spatial weighting change colour-based image segmentation?
- How well does a learned reconstruction transfer to an unseen image?
03 / PANORAMA STITCHING
Find the overlap.
Align the views.
A shared homography connected overlapping photographs to one panorama coordinate system.
Match the same scene
- I captured two overlapping views and extracted SIFT keypoints and descriptors.
- FLANN matched descriptors, while a ratio test filtered ambiguous pairs.
- RANSAC estimated a homography from the retained correspondences.
Reconstruct the image grid
- The homography mapped panorama pixels into the second photograph.
- My bilinear interpolation function combined values from the four neighbouring pixels.
- Bounds checks prevented sampling outside the input image.
Inspect the join honestly
- A narrow Gaussian filter softened the seam, while cropping reduced the output width.
- The reference-image experiment also explored brightness compensation.
- Misaligned blinds remained visible, revealing the limits of this stitching approach.
04 / IMAGE SEGMENTATION
Group the pixels.
Reveal the regions.
- I implemented K-means assignment and centroid updates using NumPy.
- Empty clusters were reinitialised to keep the iteration running.
- Adding normalised pixel positions extended colour clustering into spatially weighted superpixels.
- Lower spatial weighting favoured colour similarity across distant image regions.
- Higher spatial weighting produced more compact, location-driven groups.
- Bounding rectangles visualised cluster extents; they were not semantic object detections.
05 / IMAGE RESTORATION
Learn from examples.
Test on new images.
- A multivariate linear regression model mapped corrupted image vectors to clean image vectors.
- The implementation fitted coefficients with a pseudoinverse and reconstructed images from predictions.
- Experiments compared fixed and random corruption patterns on unseen examples.
Visual examples show reconstruction behaviour; they do not establish a verified accuracy improvement.
06 / DEVELOPMENT & ITERATION
Change a parameter.
Inspect the effect.
The work progressed from controlled transformations to more open-ended image experiments.
- 01
From known geometry to own photos
- A known homography established the coordinate and interpolation workflow.
- My own photographs then required feature matching and robust geometry estimation.
- 02
From colour to spatial structure
- Colour-only clustering established the assignment and centroid-update implementation.
- Normalised image positions introduced a controllable preference for nearby pixels.
- 03
From fixed to changing corruption
- Fixed-pattern reconstruction provided an initial image-to-image regression experiment.
- Random patterns exposed weaker reconstruction and motivated a local-patch exploration.
07 / VALIDATION & LIMITATIONS
Look at the output.
Check the assumptions.
- Saved figures exposed visible seams, spatial-weighting effects and blurred reconstructions.
- The written observations connected these effects to geometry, filtering and model generalisation.
- The current experiments provide qualitative implementation examples.
- 01
Geometry
Inspect alignment and structural discontinuities where the panorama images meet.
- 02
Segmentation
Compare colour consistency and spatial compactness as the clustering parameters change.
- 03
Reconstruction
Compare clean, corrupted and restored examples while keeping training and test images distinct.
NEXT ITERATION
Make evaluation
repeatable.
- Standardise image scaling and image selection before reporting restoration metrics.
- Re-run notebooks from a clean state to keep code and saved outputs consistent.
- Compare multiple examples before drawing conclusions about algorithm performance.
08 / REFLECTION
The lessons
I take forward.
- Implementing the calculations made image-processing assumptions tangible.
- Visual comparisons helped identify where an algorithm worked and where it needed refinement.
Features define the groups
Colour and location weighting determine which similarities the clustering algorithm treats as important.
A visual result needs a fair comparison
Consistent scaling, image selection and test data are essential before quoting performance figures.
JERRY SUN / ENGINEERING PORTFOLIO
Back to selected work