CV Image Compression Research Notebooks

This repository tracks iterative development of a segmentation-based image compression pipeline.
The project evolved from a polynomial-boundary prototype into a more robust contour-driven system using deterministic filling, hole handling, and spline-based boundary models with fallback logic.

Repository Layout

• v1.0funcCompress.ipynb to v7.3FuncCompress.ipynb: main versioned notebooks.
• v7.4funcCompress.py to v7.5FuncCompress.py: Final parameter tuning, validation, and testing
• kodak/: input images used for final testing.
• BSDS500/: BSDS500 dataset used for natural image parameter tuning and validation. It is not sent to github for space constraints
• svgs/: SVG dataset used for structured image parameter tuning, validation, and testing.
• tuned_config.json: frozen hyperparameters selected by grid search, written after tuning and read by v7.5 for all final numbers.
• requirements.txt: required packages.
• .gitignore: gitignore file.

Version Evolution

v1.0 Baseline Prototype from MPhil

• Current file: v1.0funcCompress.ipynb
• Implemented the first full pipeline: clustering, contour extraction, polynomial fitting, and reconstruction.
• Established the core serialization/reconstruction workflow.
• Main limitations at this stage were scalability and instability on larger/more complex images.

v2.0 Nearest-Neighbor and Contour Cleanup

• Current file: v2.0FuncCompress.ipynb
• Reworked cluster assignment/contour behavior to improve boundary consistency.
• Corrected issues around fitting inputs by focusing on boundary structure rather than noisy full-region point sets.
• Improved practical stability of region tracing versus the original baseline.

v2.1 Painter Ordering and Chebyshev Alternatives

• Current file: v2.1FuncCompress.ipynb
• Added painter-style rendering order (largest/background regions first, then smaller details).
• Evaluated alternatives to raw polynomial fitting (including Chebyshev-style paths) to improve numerical behavior.
• Improved reconstruction robustness where strict boundary overlap/order mattered.

v2.2 Explicit Hole-Aware Branch

• Current file: v2.2FuncCompress.ipynb
• Added explicit handling for regions with inner rings/holes.
• Introduced loop writing/parsing helpers to encode topology more directly.
• Explored trade-off between topological correctness and encoding/implementation complexity.

v3.0 Bezier Boundary Modeling and Painter's

• Current file: v3.0FuncCompress.ipynb
• Added piecewise cubic Bezier contour modeling.
• Added contour validity checks, closure controls, and improved boundary sampling logic.
• Served as a key comparison branch against polynomial and B-spline approaches.

v3.1 Closed B-Spline Transition with Painter's

• Current file: v3.1FuncCompress.ipynb
• Shifted boundary modeling toward closed B-splines.
• Improved contour smoothness and closure consistency, especially on larger images (including 256x256 cases).
• Included stronger reconstruction safeguards (coverage checks and compatibility paths during decode).

v3.2 Hole-Awareness and Chebyshev Hybrid

• Current file: v3.2FuncCompress.ipynb
• Combined explicit loop/topology handling with compact Chebyshev-style fitting.
• Focused on balancing boundary compactness with reconstruction reliability.
• Retained painter-style reconstruction logic while revisiting hole-aware behavior.

v4.0 Integrated B-Spline, Painter's, and Fallback Path

• Current file: v4.0FuncCompress.ipynb
• Shifted from HSV to CIELAB
• Uses B-spline-first boundary encoding with fallback behavior when fit quality is insufficient.
• Strengthened decode coverage handling to reduce unfilled artifacts.
• Represents the current best candidate for stable end-to-end runs.
• Implemented bilateral filter to smooth out minor color variations while strictly preserving sharp edges
• Added planar parametric color region modeling
• Implemented hybrid codec that compresses difference of original and custom algorithm with dct and reconstructs

v5.0 Ablation testing with splines vs explicit contour mapping and planar vs mean vs original color mapping

• Current file: v5.0FuncCompress.ipynb
• Ablation study implementation ☐ Exact traced contours + current region fill model: no spline fitting, fill using the exact raster contour from the segmentation stage. This removes spline error. ☐ Spline contours + current region fill model: the difference between A and B isolates boundary-model error. ☐ Exact traced contours + original per-pixel colours inside each predicted region: any remaining error in C mainly reflects mask or topology issues. ☐ Exact traced contours + a better region model: for example, per-region planar colour fit (see below) instead of mean colour. This tells you how much is due to the constant-colour interior assumption.

v5.1 Ablation testing for polyFit vs Chebyshev vs B-spline vs Bezier curves, hold planar color mapping constant

• Current file: v5.1FuncCompress.ipynb
• Ablation study implementation ☐ Polynomial.Fit ☐ Chebyshev ☐ B-Splines ☐ Bezier

v6.0 Integrated B-Spline, Painter's, and Fallback Path

• Current file: v6.0FuncCompress.ipynb
• Saved known good file, reviewed last by Dr. Town

v7.0 Completed several bug fixes, updated painter's sort algorithm, replaced fixed stride downsampling, updated segmentation similarity, region connectivity --> 8, removed redundant mean color, adaptive thresh and spline_smooth, fixed repair_coverage

• Current file: v7.0FuncCompress.ipynb
• Fixed repair_coverage function to fix bug of returning too early
• Updated spline decoding to use adaptive perimeter-based sampling (n_samples scaled by estimated contour arc length and control-point count) so large S contours are reconstructed smoothly instead of being under-sampled.
• Updated Painter’s ordering to sort decoded regions by true polygon area (cv2.contourArea) instead of contour point count, improving overlap/occlusion robustness during reconstruction.
• Replaced fixed-stride contour downsampling in spline fitting with full-contour fitting and splprep smoothing (s>0), enabling adaptive control-point allocation and better size–quality tradeoffs.
• Updated segmentation similarity to use Euclidean distance in standard-like CIELAB space (OpenCV LAB remapped to L∈[0,100], a,b* centered at 0), replacing /255 channel normalization with an approximately ΔE-interpretable threshold. (HUGE IMPROVEMENT)
• Unified region connectivity to 8-neighborhood throughout segmentation and contour extraction (CCL, border detection, and tracing) to remove diagonal-connectivity mismatches and reduce subtle boundary artifacts.
• Removed redundant mean-color storage by eliminating C records and using M (planar coefficients) as the sole region header and color model source during decoding.
• Implemented thresh (segmentation) adaptive hyperparameter: compute an image-level complexity score from local LAB variability (e.g., median local std over 5x5/7x7). Higher complexity -> lower thresh; lower complexity -> higher thresh.
• Implemented contour-adaptive spline smoothing by deriving splprep’s s per boundary from geometric complexity (perimeter, turn-angle statistics, and corner density), using lower s for sharp/corner-rich contours and higher s for smoother/longer contours.
• Replaced the previous nested-loop coverage repair with a vectorized post-fill method using scipy.ndimage.binary_dilation and convolve to identify adjacent unpainted pixels and fill them by averaging painted neighbors. Added a nearest-painted-pixel fallback (distance_transform_edt) that activates only when needed, ensuring full reconstruction coverage and removing remaining white gaps.

v7.1 Local hybrid DCT encoding and Floating to Fixed point quantization → Delta coding → Binary serialization → Entropy coding

• Current file: v7.1FuncCompress.ipynb
• Per-region DCT residual with m_k flag serialized to file
• Local rate-distortion metric for planar-only vs. planar+DCT decision (distinct from segmentation threshold)
• Replace floating-point coefficients with fixed-point/scaled integers
• Apply delta coding to coefficients before entropy coding
• Binary serialization of full record structure with actual byte counts
• Entropy coding applied to binary format (Huffman sufficient; arithmetic coding deferred to future work)
• S/L encoding ratio check per image (instrumented in encoding loop)

v7.2 Copied v5.0 and updated to match v7.1 advances and following changes

• Current file: v7.2FuncCompress.ipynb
• Add LPIPS and MSE to ablation notebook
• Align with v7.0 pipeline (8-connected segmentation, same pixel set across all six conditions)
• Bugs Bunny ablation: C vs. F results and polyline fallback percentage
• Adaptive smoothing parameter ablation
• sigmaColor x threshold interaction grid
• Acquire and prepare standard datasets (Kodak, BSDS500, structured graphics set)
• Define validation/test split; tune all hyperparameters on validation set only

v7.3 Copied v5.1 and updated to match v7.1 advances and following changes

• Current file: v7.3FuncCompress.ipynb
• Verify matched parameter budgets for B-spline vs. Bezier comparison
• Added the original polynomial fitting method as a fourth boundary baseline alongside B-spline, Bezier, and Chebyshev

v7.4 Held out set for hyperparameter tuning and validation

• Current file: v7.4FuncCompress.py
• Put everything inside functions and iterating through hyperparameter options
• Chosen parameters output to tuned_config.json which is read in by v7.5
• BSDS500: sigmaColor=350, sigmaSpace=125, high_thresh=120, low_thresh=60, smin=0, smax=5.5, bp=320.0, cw=(0.4, 0.6) (PSNR=29.68, SSIM=0.878, bpp=6.673)
• SVG: sigmaColor=85, sigmaSpace=140, high_thresh=20, low_thresh=0, smin=0, smax=15, bp=205.0, cw=(0.4, 0.6) (PSNR=28.89, SSIM=0.9329, bpp=1.191)

v7.5 Final testing numbers

• Current file: v7.5FuncCompress.py
• Consolidates v7.2 (error decomposition), v7.3 (boundary comparison), and v7.4 (codec + hyperparameter tuning) into a single evaluation harness
• Report final figures and numbers with all contributions in one file
• All figures and tables saved to files to upload to LaTeX
• v7.2/v7.3 ablation studies now run across full held-out test sets per category instead of a single image

Technical Progression Summary

1. Scalability: moved away from brittle/global behavior toward component- and contour-oriented processing.
2. Boundary modeling: progressed from raw polynomial fits to Bezier and then closed B-spline workflows.
3. Topology robustness: expanded from painter-only ordering to explicit hole-aware representations in dedicated branches.
4. Reconstruction reliability: added deterministic fill behavior, closure enforcement, and fallback/coverage safeguards.
5. Known remaining challenge: region interior appearance model is still relatively simple for texture-heavy natural images.

Serialization Notes

Note: this describes the plain-text imComp.txt format used internally by the v7.2/v7.3 ablation studies for boundary/interior comparisons. The final codec (v7.5) serializes to a separate binary format, documented in the paper.

imComp.txt record format (per region)

C;L,a,b;

Mean region color in LAB (integer values). M;xc,yc;a0,a1,a2;b0,b1,b2;c0,c1,c2;

Region centroid (xc, yc) followed by planar LAB model coefficients, one triple per channel. Model: value(x,y) = a + b*(x - xc) + c*(y - yc). S;k;knots;ctrl_row;ctrl_col;

Spline boundary record. k: spline degree (integer). knots: comma-separated knot vector. ctrl_row: comma-separated row control points. ctrl_col: comma-separated column control points. L;r,c;r,c;...;

Fallback loop/polyline boundary record. Sequence of sampled contour points as row/column pairs. Parsing and grouping rules

Fields are separated by semicolons (;). Numeric tuples/lists inside a field are comma-separated (,). A region is stored as: C + M + (S or L) followed by a blank line. S is preferred; L is used when spline fitting fails.