--- title: 'ImPASS: Image Phase Alignment Super-Sampling' url: https://www.emergentmind.com/topics/image-phase-alignment-super-sampling-impass type: topic --- # ImPASS: Image Phase Alignment Super-Sampling Searching arXiv for ImPASS and closely related papers to ground the article in current literature. Image Phase Alignment Super-Sampling (ImPASS) is a computational imaging algorithm for converting a sequence of displaced low-resolution images into a single high-resolution image. In the reported implementations, it combines Phase Correlation image registration with SeDDaRA blind deconvolution, and has been validated in simulations, in laboratory imaging, and in widefield microscopy. In the first microscopy study, 80-frame image sets were up-sampled by a factor of eight, aligned, combined, and processed; the reported measurement showed a resolution improvement by a factor of 2.68 and a final resolution 1.79 below the diffraction limit. Subsequent algorithmic modifications further increased the reported improvement to as much as 4.41 relative to the original image resolution and 2.57 below the diffraction limit [2508.21714; 2601.04391]. ## 1. Definition and operating regime ImPASS is designed to synthesize a single high-resolution image from a sequence of low-resolution images of a static scene, where each frame contains sub-pixel shifts. The fundamental steps are measurement of relative displacements, up-sampling, alignment, combination, and blind deconvolution. The method is explicitly multi-frame: the input is a sequence of \(N\) low-resolution frames, each displaced by a fraction of a pixel in a controlled or measured fashion, and the displacement grid is designed so that, in combination, the sequence provides dense sampling across the super-resolved grid [2508.21714]. The method targets regimes in which the native sensor sampling is insufficient and, in the reported microscopy experiments, where the optical diffraction limit is also a practical benchmark. In the widefield microscopy application, no customization beyond a translation stage was required. The acquisition used passively illuminated sequences and standard widefield hardware, rather than specialized illumination or phase-sensitive detection [2508.21714]. This operating regime distinguishes ImPASS from single-image super-resolution pipelines and from methods that require active optical encoding. A plausible implication is that its performance depends jointly on sampling geometry, registration accuracy, and deconvolution, rather than on any single component in isolation. ## 2. Registration, up-sampling, and image combination The core registration mechanism in ImPASS is phase correlation. For a current image \(I_c(m,n)\) and a target image \(I_t(m,n)\) related by a sub-pixel translation \((\Delta x,\Delta y)\), \[ I_t(m,n)=I_c(m-\Delta x,n-\Delta y). \] After 2D Fourier transformation, the translation appears as a linear phase shift, \[ F_t(u,v)=e^{-i(u\Delta x+v\Delta y)}F_c(u,v), \] and the phase difference is written as \[ e^{i(u\Delta x+v\Delta y)}= \frac{F_c(u,v)F_t^*(u,v)}{|F_c(u,v)F_t^*(u,v)|}. \] The inverse transform of this phase difference yields a correlation peak at \((\Delta x,\Delta y)\), and centroid analysis of the peak yields sub-pixel accuracy. This registration method is described as robust to noise and efficient compared to other registration methods [2508.21714]. After displacement estimation, each low-resolution frame is interpolated to a higher-resolution grid. In the microscopy study, cubic convolution interpolation with parameter \(-0.5\), as per Park and Schowengerdt, was used, although the exact parameter was noted as not critical. The reported up-sampling factor was \(M_{SS}=8\), so that each original pixel was mapped to an \(8\times 8\) higher-resolution grid. Each up-sampled image was then shifted according to the measured displacement vectors, and the aligned sequence was combined by simple pixel-wise averaging [2508.21714]. The averaged image is not itself the final result. The combined image typically appears blurry, because it remains limited by the system point spread function (PSF). ImPASS therefore treats registration and averaging as a super-sampling stage that densifies the sampling lattice before a subsequent deblurring stage recovers higher-frequency structure. ## 3. SeDDaRA blind deconvolution and frequency recovery The deconvolution stage in ImPASS is performed by SeDDaRA blind deconvolution. In the microscopy report, SeDDaRA is described as a non-iterative blind deconvolution technique that extracts the PSF from the image by comparing the spatial-frequency content of the blurred combined image with that of a reference distribution, where the reference need only possess appropriate spatial-frequency characteristics and need not have similar appearance. The stated purpose of this stage is to recover high-frequency details and undo the blurring imposed by the system PSF after multi-frame combination [2508.21714]. The same report states that SeDDaRA allows recovery of spatial-frequency components outside the passband of the imaging system, namely frequencies lost to diffraction or sensor limitations. In the microscopy experiments, SeDDaRA was applied with the radius of influence set to 24 pixels and the regularization parameter \(C_2=0.02\), balancing frequency recovery against noise and artifact amplification [2508.21714]. The later modification paper expresses the processed image in Fourier space as \[ F(u,v)=\frac{G(u,v)D^*(u,v)}{|D(u,v)|^2+C_2D_{ave}}, \] where \(G(u,v)\) is the Fourier transform of the combined image, \(D(u,v)\) is the extracted PSF, \(D_{ave}\) is the average of \(|D(u,v)|\), and \(C_2\) is the regularization parameter. In the modified algorithm, the denominator treatment was changed so that values of \(|D(u,v)|^2\) below \(C_2D_{ave}\) were set to the threshold value rather than having \(C_2D_{ave}\) added uniformly; this was reported to further limit noise amplification in low-signal-frequency regions [2601.04391]. Taken together, these stages define the characteristic ImPASS workflow: acquisition of displaced frames, phase-correlation registration, high-factor up-sampling, alignment, averaging, and blind deconvolution. ## 4. Widefield microscopy realization The first widefield microscopy application of ImPASS used an Olympus Inverted Fluorescence Microscope IX-83 with an Olympus UPlanFL 4x lens of numerical aperture \(NA=0.13\), a Hammatsu ORCA-Flash 4.0 v3 camera with \(6.5\,\mu\mathrm{m}\times 6.5\,\mu\mathrm{m}\) pixels, an Olympus IX3-SSU translation stage providing sub-micron precision, and broadband visible light illumination. The samples were a 5\(\mu\)m H\&E-stained Porcine Cornea slice for qualitative assessment and a standard US Air Force resolution chart for quantitative assessment [2508.21714]. Each sample sequence contained 80 or 81 frames. The stage displacement was approximately 990 nm per step along the diagonal in an “X” pattern. Original images of \(2034\times 2042\) pixels were cropped to \(532\times 532\) for processing, and after up-sampling by a factor of 8 the cropped images became \(4256\times 4256\). Registration was implemented by a Python/IDL-based script using phase correlation, because translation stage readings alone were verified to be imprecise; interpolation was performed in IDL, and SeDDaRA deconvolution was performed in Quarktet Tria software [2508.21714]. Quantitative resolution assessment used both line-pair counting and the formal slant-edge method of ISO 12233. In the slant-edge procedure, a slanted edge is used to build an edge spread function, a sigmoid is fit, and the derivative—the line spread function or PSF—yields the effective resolution through its full width at half maximum (FWHM). The optical benchmark was the Abbe criterion, \[ d=\frac{\lambda}{2NA}. \] For \(NA=0.13\) and \(\lambda=650\,\mathrm{nm}\), the reported diffraction limit was \(2.5\,\mu\mathrm{m}\) [2508.21714]. The reported microscopy results were as follows. For the Porcine cornea sample, features not resolved in the original 4x images became clear in the super-resolved result, and the processed image was described as comparable to or better than a “truth” image taken with a 10x objective, suggesting at least a 2.5x resolution improvement. For the USAF chart, the original image resolved group 7, element 6, corresponding to 228 line pairs/mm, whereas the ImPASS result resolved group 8, element 4, corresponding to 362 line pairs/mm. By the slant-edge metric, the pre-processed FWHM was \(3.73\,\mu\mathrm{m}\), the post-processed FWHM was \(1.39\,\mu\mathrm{m}\), the resolution improvement factor was 2.68, and the final resolution lay 1.79 below the diffraction limit of \(2.5\,\mu\mathrm{m}\) [2508.21714]. This suggests that the reported sub-diffraction behavior is a property of the combined acquisition-and-reconstruction pipeline rather than of the native widefield optics in isolation. ## 5. Algorithmic modifications and later performance The 2026 modification study retained the basic ImPASS structure—measurement of relative displacements, up-sampling, alignment, combination, and blind deconvolution—but altered both alignment and deconvolution to improve resolution [2601.04391]. For sub-pixel peak localization in phase correlation, the modified method introduced a generalized centroid: \[ \Delta x= \frac{\sum_{m=m_p-a}^{m_p+a} m\,[I_{cp}(m,n)]^p} {\sum_{m=m_p-a}^{m_p+a}[I_{cp}(m,n)]^p}, \qquad \Delta y= \frac{\sum_{n=n_p-a}^{n_p+a} n\,[I_{cp}(m,n)]^p} {\sum_{n=n_p-a}^{n_p+a}[I_{cp}(m,n)]^p}, \] where \(I_{cp}(m,n)\) is the intensity at the correlation peak, \((m_p,n_p)\) is the pixel containing the maximum, \(a\) is the window half-width, and \(p\) is the centroid power. Previous ImPASS used \(p=1\), whereas the modification paper reports that \(p=2\), described as a “centroid of squares,” minimized registration error, reducing misalignment by up to 50% in one axis and 17% in the other on test data [2601.04391]. A second modification was multi-pass phase correlation. Instead of aligning all frames to a single reference, the method cycles each image as the control, producing an alignment displacement matrix \(D(c,z)\). The reported normalization and averaging are \[ D_{new}(c,z)=D(c,z)-D(c,0), \qquad D_{fin}(z)=\frac{1}{Q}\sum_{c=0}^{Q-1}D_{new}(c,z). \] This reduced systematic alignment errors, particularly for large fractional-pixel shifts, and the paper reports error reduction by factors up to 3.11 and 1.44 on two test sets [2601.04391]. The third modification was the SeDDaRA denominator threshold already noted above. Rather than adding \(C_2D_{ave}\) to all denominator values, the threshold was applied only where \(|D(u,v)|^2