Source-linked AI summary
A New Algorithm For Difference Image Analysis
D. M. Bramich
TL;DR
DIA requires a convolution kernel to match images for accurate difference photometry, but basis-function models require user-selected parameterizations and can be sensitive to alignment. The paper directly solves kernel pixels by linear least-squares, extending the method across image sub-regions; author-reported tests show better difference images and support misaligned or trailed images.
Problem
Existing DIA kernel models require users to choose basis functions and related parameters, creating a complex, dataset-dependent fitting problem.
Method
The method represents the convolution kernel as a discrete pixel array, solves its pixel values by linear least-squares, and extends spatial variation through grid-based sub-region solutions.
Results
The method produced higher-quality difference images than ISIS2.2 and handled image shifts, telescope jumps, and trailed images within the reported tests.
Takeaways & Limitations
The flexible kernel model supports image matching without resampling for simple translations and can reduce data loss from tracking or focus errors.
Takeaways & Limitations
The reference image should ideally be the better-seeing, high-signal-to-noise image, often formed by stacking best-seeing images.
Abstract
from arXiv · showhide
In the context of difference image analysis (DIA), we present a new method for determining the convolution kernel matching a pair of images of the same field. Unlike the standard DIA technique which involves modelling the kernel as a linear combination of basis functions, we consider the kernel as a discrete pixel array and solve for the kernel pixel values directly using linear least-squares. The removal of basis functions from the kernel model is advantageous for a number of compelling reasons. Firstly, it removes the need for the user to specify such functions, which makes for a much simpler user application and avoids the risk of an inappropriate choice. Secondly, basis functions are constructed around the origin of the kernel coordinate system, which requires that the two images are perfectly aligned for an optimal result. The pixel kernel model is sufficiently flexible to correct for image misalignments, and in the case of a simple translation between images, image resampling becomes unnecessary. Our new algorithm can be extended to spatially varying kernels by solving for individual pixel kernels in a grid of image sub-regions and interpolating the solutions to obtain the kernel at any one pixel.
1 INTRODUCTION
DIA matches images by deriving a convolution kernel and enables accurate time-series photometry, especially in crowded fields. The paper proposes a simpler, more robust kernel-solving algorithm than the current Alard approach.
- DIA derives a convolution kernel describing point-spread-function changes between images.
- DIA can achieve errors near theoretical Poisson limits and is reliable for exceptionally crowded stellar fields.
- The Alard algorithm determines kernels in image space by decomposing them into basis functions and supports spatially varying solutions.
- The proposed algorithm retains linear least-squares while requiring fewer input parameters and offering simpler implementation and greater robustness.
2 A NEW APPROACH TO THE KERNEL SOLUTION
The new approach replaces user-selected kernel basis functions with a directly solved pixel-array kernel and differential background. It uses weighted linear least-squares while accommodating practical image defects and producing difference images for photometric analysis.
- Motivation: The Alard method requires users to choose Gaussian basis functions, widths, polynomial degrees, and complexity appropriate to each dataset.This parameterization can require substantial experimentation because different situations may need different basis combinations.
- Pixel-array kernel: The new method represents a spatially invariant convolution kernel as a pixel array and solves its pixel values with an unknown constant differential background.
- Least-squares solution: Least-squares minimization produces a linear system whose unknowns are the kernel pixels and differential background.The solution is obtained by constructing U and b and calculating a = U^-1b.
- Practical handling: Bad or saturated pixels are excluded, but one defective reference pixel can invalidate current-image pixels across the kernel area.The method therefore recommends minimizing bad reference pixels and avoiding excessively large kernels.
- Photometric scaling: The kernel-pixel sum is the photometric scale factor, incorporating relative exposure time and atmospheric extinction.Subtracting a constant reference-image background can reduce correlation between this factor and the fitted differential background.
- Difference image: Difference images subtract the fitted model from the current image, leaving variable-source or changing-background residuals for difference-flux measurement.
- Iterative fitting: The kernel and background require iteration because pixel variances depend on the model values.
A New Algorithm For Difference Image Analysis 3
The algorithm estimates uncertainties from a CCD noise model and iteratively fits the image model while clipping outliers. It extends to spatially varying kernels by solving on a grid and interpolating between sub-regions.
- Uncertainty model: The uncertainty model uses CCD readout noise, gain, and a master flat field to calculate pixel variances.The reference and master-flat images are treated as noiseless high-signal-to-noise inputs.
- Iterative fitting: The fitting procedure initializes the model with current-image values, updates variances iteratively, and applies 3σ clipping to exclude outliers.
- Spatial variation: Spatially varying solutions divide the image into an Nx by Ny grid, solve each sub-region, and interpolate kernel and background solutions at any image pixel.
3 COMPARISONS WITH THE ALARD ALGORITHM
The tests show that DANDIA’s pixel-based kernel solution handles translations, resampling, and complicated kernel shapes more accurately than ISIS2.2’s basis-function approach, while O-MOMS improves resampling when it is used.
- Initial tests: DANDIA exactly recovers a one-pixel translation kernel, whereas ISIS2.2 leaves substantially worse residuals with SISIS/SDANDIA ≈26190.The test uses an unresampled shift by one pixel in both spatial directions.
- Initial tests: Both methods recover a Gaussian blur with FWHM 4.0 pix, but DANDIA still outperforms ISIS2.2 with SISIS/SDANDIA ≈1510.The Gaussian case is the test in which ISIS2.2 can closely model the kernel.
- Initial tests: DANDIA nearly perfectly recovers the complicated half-pixel-shift kernel, while ISIS2.2 fails, yielding SISIS/SDANDIA ≈18450.The half-pixel shift requires resampling, performed here with cubic O-MOMS.
- Initial tests: ISIS2.2 is limited to kernels near the centre with scales matching its Gaussian sigmas, whereas DANDIA imposes no kernel-shape assumption.The comparison attributes DANDIA’s broader performance to its direct pixel-array kernel representation.
- Image resampling: Cubic O-MOMS resampling is used because it has high approximation order and a small approximation-error constant for a given support.The method is presented as an improvement when image resampling is employed.
- Spatially varying kernels: The spatially varying DANDIA solution uses kernels on a 10×10 grid and bilinear interpolation to obtain the kernel at each image pixel.Each circular kernel contains 317 pixels.
A New Algorithm For Difference Image Analysis 5
Tests E–H show that DANDIA produces better difference images than ISIS2.2, especially when images are aligned without resampling. The chi-squared analysis detects these differences even when the normalized images look nearly identical.
- Integer-pixel registration: Without resampling, DANDIA achieved χ2/NDOF = 0.99 and ISIS2.2 achieved 1.00.Tests G and H registered the images using an integer-pixel translation.
- Algorithm comparison: DANDIA outperformed ISIS2.2 by 0.60σ with resampling and by 0.38σ without resampling.These comparisons are based on the χ2 analysis rather than visual inspection.
- Effect of resampling: Image resampling degraded the difference images by 0.48σ for DANDIA and 0.70σ for ISIS2.2.The strongest result was test G, using DANDIA without resampling, which performed 1.08σ better than test F.
- Visual versus statistical assessment: The normalized difference-image cutouts appeared very similar, with visible differences emerging only under detailed scrutiny.The χ2 analysis nevertheless distinguished the quality of the four tests.
4 CONCLUSIONS
The paper introduces a pixel-array convolution-kernel model that avoids user-selected basis functions and can incorporate image translations directly. The method produces higher-quality difference images and supports applications where conventional methods fail.
- Kernel model: The new method models the convolution kernel as a pixel array rather than a combination of user-selected basis functions.This removes the need to specify basis functions and avoids potentially inappropriate choices.
- Image registration: For translated images, the kernel incorporates resampling directly, avoiding image resampling and its correlated pixel noise.Basis-function kernels may only partly compensate for sub-pixel translations because their functions are centered at the kernel-coordinate origin.
- Applications and consequences: The method produced higher-quality difference images than ISIS2.2 and can construct images for telescope jumps or trailed images, where ISIS2.2 fails.The authors connect improved difference-image quality with more accurate light curves and potentially less data loss from tracking or focus errors.