shinka 0.3.0

A high-quality image upscaling package using a genetically-evolved algorithm.

SHINKA Upscaler

SHINKA is a high-quality, non-ML image upscaling package for Python, built around a custom, genetically-evolved 5-tap convolution kernel. The kernel, nicknamed the halo tap, was discovered through large-scale GPU-accelerated search and frequency-domain optimization—not by maximizing SSIM or any perceptual metric, but by directly optimizing the frequency response for passband fidelity and stopband suppression. The result is a lightweight, dependency-minimal upscaler that delivers state-of-the-art perceptual quality for both grayscale and color images.

Note: SHINKA was not designed to maximize SSIM or any learned perceptual metric. Its “halo tap” kernel is the result of a pure frequency-domain search for the best trade-off between sharpness and anti-aliasing.

What Makes SHINKA Unique?

• Algorithmic, Not ML: No neural networks, no training data, no model weights. SHINKA's core is a 5-tap symmetric kernel, evolved to maximize frequency-domain performance.
• The “Halo Tap”: The final kernel is a numerical twin of the original “halo tap” discovered via computational search.
• Genetic & Gradient Search: The kernel was discovered by running millions of random and gradient-optimized candidates on GPU, using a frequency-domain objective that balances pass-band fidelity and stop-band suppression. See discovery/conv_1d.py for the search code.
• Fast & Lightweight: Pure Python (with NumPy, SciPy, Pillow), no heavy dependencies. Suitable for both scripts and production pipelines.
• CLI & API: Use as a Python library or from the command line for batch or single-image upscaling.

Features

• Upscales images by an integer factor (e.g., 2x, 4x)
• Preserves color and ICC profile
• CLI for batch and single-image upscaling
• Python API for integration in scripts and pipelines

Performance Metrics

• Passband MSE: 1.94 × 10⁻²
• Stopband Peak: 0.57 (normalized, cf. Lanczos-3: 1.0)
• Computational Cost: 5 multiplications/additions per output pixel
• Pareto-Optimal: No standard kernel (bilinear, symmetric 5-tap, Lanczos-7/9) dominates this kernel across all metrics.
• Visual and Analytical Tests: Superior suppression of high-frequency noise and preservation of image features compared to standard alternatives.

Installation

Install the core upscaler:

pip install shinka

For CLI usage:

pip install shinka[cli]

Python API Usage

from shinka import upscale

# Upscale an image from file and save the result
result_path = upscale("input.jpg", scale=2, save_path="upscaled.png")

# Upscale a NumPy array and get a PIL Image
import numpy as np
from PIL import Image
arr = np.array(Image.open("input.jpg"))
result_img = upscale(arr, scale=2)  # returns PIL Image

Command-Line Usage

After installing with [cli], you get a shinka command:

shinka single input.jpg --scale 2 --output upscaled.png
shinka batch ./images --scale 2 --output ./upscaled

• single: Upscale a single image file
• batch: Recursively upscale all images in a directory (supports PNG, JPEG, etc.)

Algorithm Details

• Kernel: 5-tap symmetric, two-phase optimized (“halo tap”)
• Convolution: Separable 1D, applied in both axes
• Boundary Modes: Reflect, constant, nearest, mirror, wrap
• No ML/AI: No training, no model weights, no external data

---

Technical Appendix: Kernel Discovery and Analysis

Title

A Pareto-Optimal 5-Tap Symmetric Kernel for Efficient Image Upscaling: Discovery via Large-Scale Random Search and Gradient-Based Refinement

Abstract

We present a computationally efficient, Pareto-optimal 5-tap symmetric convolution kernel for image upscaling, designed to maximize passband fidelity and stopband suppression while minimizing computational cost. The kernel was discovered through a two-stage process: (1) a large-scale, GPU-accelerated random search over the space of normalized symmetric 5-tap filters, and (2) gradient-based fine-tuning using frequency-domain objectives. This approach yields a kernel that outperforms standard alternatives (e.g., bilinear, Lanczos-3/7) in the trade-off between computational efficiency and anti-aliasing performance. We discuss the methodology, analyze the resulting kernel’s properties, and outline the generalizability of this approach to other kernel design problems.

1. Introduction

Image upscaling is a fundamental operation in computer vision, graphics, and imaging pipelines. The choice of interpolation kernel directly impacts both visual quality and computational efficiency. While high-tap kernels (e.g., Lanczos-7) offer strong anti-aliasing, they are computationally expensive. Conversely, low-tap kernels (e.g., bilinear) are efficient but suffer from poor frequency response and aliasing artifacts. There is a practical need for kernels that optimally balance these competing objectives.

2. Methodology

2.1. Search Space

We restrict our search to symmetric, normalized 5-tap kernels, parameterized as:

$$ h = [a, b, c, b, a], \quad \text{with} \quad \sum h = 1 $$

where $a, b, c$ are real-valued parameters.

2.2. Objective Function

The kernel’s frequency response is evaluated using the discrete Fourier transform. The objective function is:

$$ J = \text{MSE}_{\text{passband}} + \alpha \cdot \max(\text{stopband}) $$

where $\text{MSE}_{\text{passband}}$ measures deviation from the ideal (unity) response in the passband, and $\max(\text{stopband})$ penalizes the largest response in the stopband. The weighting parameter $\alpha$ controls the trade-off.

2.3. Large-Scale Random Search

We employ a GPU-accelerated random search, generating hundreds of millions of candidate kernels, each evaluated for frequency-domain performance. The best candidate is selected as the starting point for refinement.

2.4. Gradient-Based Fine-Tuning

The selected kernel is further optimized using gradient descent (Adam optimizer) on the objective $J$, with staged adjustment of $\alpha$ to first flatten the passband and then tighten the stopband. This two-phase approach ensures both high fidelity and strong anti-aliasing.

3. Results

The discovered 5-tap kernel achieves:

• Passband MSE: $1.94 \times 10^{-2}$
• Stopband Peak: $0.57$ (normalized, cf. Lanczos-3: 1.0)
• Computational Cost: 5 multiplications/additions per output pixel

Pareto analysis demonstrates that no standard kernel (bilinear, symmetric 5-tap, Lanczos-7/9) dominates this kernel across all metrics. Visual and analytical tests confirm superior suppression of high-frequency noise and preservation of image features.

4. Discussion

4.1. Importance of the 5-Tap Kernel

This kernel fills a critical gap in the design space: it offers strong anti-aliasing and high fidelity at a computational cost suitable for real-time and embedded applications. Its symmetric, short support makes it ideal for separable 2D filtering and hardware implementation.

4.2. Generalizability

The methodology—combining large-scale random search with gradient-based refinement—can be generalized to other kernel design problems:

• Different tap counts: The approach scales to 3-tap, 7-tap, or higher-order kernels.
• Asymmetric or constrained kernels: The parameterization and objective can be adapted for non-symmetric or application-specific constraints.
• Other domains: The same framework applies to audio, 1D/3D signals, or any domain where frequency-domain kernel properties are critical.

4.3. Future Applications

Potential applications include:

• Real-time video upscaling in resource-constrained environments
• Hardware-accelerated image processing pipelines
• Custom anti-aliasing filters for scientific imaging or medical devices
• Audio resampling with strict computational budgets

5. Conclusion

We have demonstrated that a data-driven, computational search approach can yield a 5-tap upscaling kernel that is Pareto-optimal and practically superior to standard alternatives. This methodology is broadly applicable to kernel design problems where trade-offs between fidelity, anti-aliasing, and efficiency are paramount.

References

• Oppenheim, A. V., & Schafer, R. W. (2009). Discrete-Time Signal Processing. Pearson.
• Smith, J. O. (2007). Introduction to Digital Filters: With Audio Applications. W3K Publishing.
• Harris, F. J. (1978). On the use of windows for harmonic analysis with the discrete Fourier transform. Proceedings of the IEEE, 66(1), 51-83.
• [conv_1d.py source code, this work]

---

License

This project is licensed under the MIT License. See the LICENSE file for details.