A new computer-vision technique could make hidden vibrations and subtle shape changes visible in objects whose boundaries are too complex for conventional tracking methods. The approach, developed by researchers Qiang Tang, Mengjing Shi, Yingying Chen, Jie Guo and Dashan Zhang, replaces idealized geometric models with a reference boundary constructed directly from image data. In tests involving a vibrating PVC pipe and a stretched polymer lattice, the method amplified tiny departures from an object’s baseline shape while preserving the broader structure of the image. The results, reported in Results in Optics, point toward a more flexible way to monitor mechanical deformation without attaching markers, reflective targets or specialized sensors to the structure.
Measuring vibration from video is attractive because cameras can observe an entire object from a distance and capture spatially distributed motion rather than readings at only a few sensor locations. Yet many established vision-based techniques depend on texture, speckle patterns or clearly identifiable markers. Optical flow, image registration and digital image correlation all become more difficult when an object has a smooth, repetitive or weakly textured surface. For edge-dominated structures, the most informative signal may not be texture movement at all, but the minute displacement of a boundary. A pipe wall that flexes, a mesh opening that elongates or a thin edge that develops a local dent can reveal deformation through geometry even when its surface contains little trackable detail.
The new framework builds on a concept known as deviation magnification. Unlike Eulerian video magnification, which amplifies temporal changes in image intensity or phase at fixed pixel locations, deviation magnification treats deformation as a geometric departure from a reference shape. A reference edge is established, and the observed boundary is sampled in directions perpendicular to that edge. Small normal displacements are then estimated as a one-dimensional signal along the boundary. Once amplified, those displacements can be used to warp the image, making otherwise imperceptible shape changes visible. The critical problem is that the reference edge must be accurate: if it is poorly fitted, the method may magnify errors in the model rather than genuine structural motion.
Earlier implementations commonly used simple parametric references such as straight lines, circles or ellipses. These models work well when an object really does resemble one of those shapes, but they can fail when curvature changes sharply, when local features are important or when the boundary has an arbitrary form. A low-order model may cut across a complex contour, causing the normal sampling region to miss part of the real edge. A highly flexible model may follow noise or local irregularities too closely. Because neither the perfectly undeformed boundary nor the true deformation is directly visible in a real image, the researchers sought a compromise: a reference that stays close to the detected edge while smoothing away short-scale fluctuations.
Their solution uses a technique called principal-curve estimation by local means, or PCLM. The process begins with detected edge points that have been arranged in order along a single boundary. For each point, the algorithm examines a neighborhood and calculates the local mean position and a two-dimensional covariance matrix. The dominant eigenvector of that matrix identifies the prevailing tangent direction of the boundary, while the perpendicular vector defines a local normal. The current reference point is then projected onto the locally estimated tangent line. Repeating this operation acts like a geometric low-pass filter: small deviations are suppressed, but the large-scale course of the edge is retained. In the study, five iterations were sufficient because the detected edge was already close to the desired reference.
The size of each neighborhood determines how aggressively the boundary is smoothed, making its automatic selection central to the method. Small neighborhoods keep the reference close to local features but may preserve noise and produce an undulating baseline. Large neighborhoods create a smoother curve but can flatten legitimate geometry or pull the reference inward, introducing a low-frequency bias into the measured deviation. To choose between candidate neighborhood sizes, the researchers scored each resulting reference using two measures of closeness and one measure of smoothness. The global residual penalizes systematic drift, while a robust local residual detects sections where the reference mismatches the edge. A discrete bending energy favors a geometrically regular curve. After normalizing these criteria, the algorithm selects the neighborhood that minimizes the worse of the global and local mismatch, plus a weighted smoothness penalty.
The image-processing pipeline first identifies subpixel edge locations using a partial-area method, then groups nearby detections with density-based clustering. Isolated points and very small clusters are discarded as noise. Within each remaining cluster, a nearest-neighbor travelling-salesman heuristic orders the points along the boundary, allowing the method to distinguish open curves from closed contours. Once a reference curve has been constructed, the algorithm samples image intensity along its local normals. At each position, the observed intensity profile is compared with a common profile formed by averaging all normal profiles along the edge. Assuming that local profiles differ mainly by translation, a first-order least-squares estimate extracts the normal displacement. In simplified form, the estimate compares the difference between a local profile and the average profile with the derivative of the average profile, allowing the displacement to be recovered even when it is smaller than a pixel.
The researchers tested the method first on synthetic boundaries, where the true reference and prescribed deviations were known. These tests demonstrated how strongly reference selection affects the result. For one curved structure, the adaptive procedure selected a neighborhood of 36 points and produced a root-mean-square error of 0.532, close to the posterior best value of 0.529 obtained using a known reference–deviation decomposition. For a near-linear structure, the algorithm detected that the boundary was effectively straight using a line-likeness statistic. When the statistic exceeded 0.99, the method bypassed the curved-reference optimization and used all edge points to construct a global straight-line fit. In the synthetic near-linear case, this produced an RMSE of 0.222. The special case also reproduced the behavior of the original straight-line deviation-magnification approach, providing a check that the generalized framework did not disrupt simpler geometries.
The most visually striking test involved a PVC pipe struck by a hammer. The experiment used high-speed video recorded at 24,096 frames per second with a 192-by-192-pixel resolution. From a region of interest, the algorithm identified 871 subpixel edge points and separated them into two closed contours representing the pipe’s outer and inner boundaries. Because neither contour was close to linear, PCLM constructed separate smooth reference curves, automatically selecting neighborhood sizes of five points for the outer contour and four for the inner contour. Eleven intensity samples were taken along each normal. The reference was fixed from an initial frame, and subsequent frames were analyzed relative to it. Before the impact, the extracted deviations mainly represented static geometric differences between the pipe and its surrogate baseline. After the hammer strike, they captured the transient response as well. The magnified images revealed the localized dent at the impact point and its later rebound, showing how spatial deformation and its evolution through time can be visualized together.
A second experiment pushed the method toward a more difficult geometry: a square polyethylene mesh that was stretched vertically. Unlike the pipe, the mesh contained many short segments and repeating openings, creating a dense network of boundaries rather than two smooth contours. In a 430-by-280-pixel region of interest, subpixel detection found 2,982 edge points, which were grouped and ordered into 18 separate edge structures. The system built a data-driven reference for each structure and generated local normals for sampling. With the reference configuration fixed at the first, pre-stretched time point, the deviation fields showed how deformation varied across the mesh as stretching continued. Some edge groups and cells shifted more strongly than others, revealing heterogeneous deformation. The resulting magnified images made the progressive elongation and shape changes of the openings visible while retaining the overall connectivity of the lattice.
The method is not a universal substitute for physical sensors, and the authors emphasize a fundamental limitation shared by reference-based deformation analysis. An exact undeformed boundary cannot be recovered unambiguously from a single image. Even a data-driven reference remains a surrogate, so the residual between that curve and the observed edge may contain modeling error. The distinction matters because magnifying the wrong component can generate spectacular but physically meaningless distortions. In a single pipe frame with no intended motion, magnifying the estimated deviation produced a modest and plausible enhancement. Magnifying the static geometric offset between the detected edge and the reference by a factor of 500, however, created a severe artificial warp. The experiment illustrates why reference construction and deviation separation are as important as the final visual amplification.
The researchers say the framework extends deviation magnification from simple ideal shapes to complex and arbitrary boundaries, with applications potentially spanning structural monitoring, laboratory mechanics, vibration analysis and full-field deformation measurement. Its principal advantage is that it requires no predefined global equation for the object’s shape. Instead, the reference adapts to the boundary observed in the image, while the neighborhood-selection rule balances fidelity against smoothness without needing the ground truth. The method still depends on reliable edge detection and can be challenged by poor lighting, occlusion, clutter, severe noise or deformations that violate the assumption of a locally smooth curve. Future improvements will focus on making parameter selection more robust and extending the approach to more demanding imaging conditions. For now, the experiments suggest that a camera can expose hidden mechanical behavior not by tracking texture, but by carefully reconstructing—and then magnifying—the geometry of an object’s edge.
Subject of Research: Data-driven video-based deviation magnification for measuring subtle deformation and vibration along complex object edges.
Subject of Research: Technology and Engineering
Article Title: Deviation magnification of complex edge structures via adaptive reference construction
Article References: Tang, Q., Shi, M., Chen, Y., Guo, J., & Zhang, D. (2026). Deviation magnification of complex edge structures via adaptive reference construction. Results in Optics, Article 101124. https://doi.org/10.1016/j.rio.2026.101124
Image Credits: AI Generated
DOI: 10.1016/j.rio.2026.101124
Keywords: deviation magnification, computer vision, structural vibration, edge detection, video-based measurement, principal curves, image warping, deformation analysis
Cite this news
APA MLA Chicago
SCIENMAG. (August 28, 2026). Adaptive References Magnify Deviations in Complex Edge Structures. https://scienmag.com/adaptive-references-magnify-deviations-in-complex-edge-structures/
Copy citation Download RIS
Tags: boundary deviation detectionboundary-based vibration measurementcomplex edge boundary detectioncomplex object boundary trackingcomputer visioncomputer vision for mechanical deformationdeformable object boundary analysisedge structure analysisimage data-driven shape trackingimage-based mechanical deformation sensingimage-based structural health monitoringmarkerless deformation analysismarkerless vibration measurementnon-contact structural health monitoringnon-contact vibration monitoring techniquesoptical flow limitationsoptical flow limitations in smooth surfacesreference boundary construction from image datashape deformation monitoringshape deviations amplification methodssubtle shape change amplificationsubtle shape change visualizationvisualization of tiny boundary deviations
