Area-Driven Adaptive Sampling of Closed Droplet Contours for Vision-Based Droplet Observation Authors: Xuefeng Wang, Yangting Zheng, Chenyao Bai, Yinqi Chen, Xiang Gao, Yiyue Li, Yunlong Zhu Published: 2026 ABSTRACT Closed droplet contours provide the geometric basis for area estimation in vision-based droplet observation. In OLED inkjet printing, droplets are deposited into pixel wells with predefined geometry; projected area is therefore a primary geometric quantity for assessing whether the deposited liquid sufficiently fills the well or risks overflow. This work formulates closed-contour sampling under a fixed sampling budget as an area-driven sampling problem. A leading-order analysis of the local arc–chord area error shows that the dominant cubic term depends jointly on curvature and segment length. Minimization of the resulting leading-order area-error functional yields an asymptotically optimal area-driven sampling density proportional to the cube root of curvature, together with a sampling-budget estimate under a target area-error tolerance. The derived sampling density is implemented on the fitted closed contour through cumulative-weight inversion. Experiments on random closed curves and the droplet dataset provide a systematic quantitative comparison with representative methods under identical fixed-budget settings, complemented by statistical analysis and evaluations of geometric fidelity, sensitivity, and computational efficiency. The proposed method achieves lower area estimation error under the tested sampling budgets, with the improvement being most pronounced at lower sampling budgets, while the reported geometric-fidelity metrics show no disproportionate degradation of contour fidelity. These results demonstrate the effectiveness of area-driven sampling for closed-contour area estimation under limited sampling budgets. 1. Introduction In vision-based droplet observation, images serve as the geometric foundation for a range of downstream tasks, including projected-area estimation, volume inversion, morphology-stability analysis, and process-parameter optimization. Among all observable features, the closed contour extracted from droplet images provides the most direct and stable link between the image domain and physically meaningful quantities [1,2,3], such as projected area, equivalent diameter, roundness, and eccentricity. These quantities hold direct engineering importance in applications like inkjet printing, liquid dispensing, and spray systems [4,5,6,7]. Nevertheless, for a closed contour with an arbitrary shape, its enclosed area is ultimately evaluated after discretizing the contour into an ordered set of boundary points, which are represented by a polygonal approximation. As a result, the area computed from a discretely sampled polygon inevitably incurs approximation errors (Figure 1). This observation leads directly to the central problem addressed in this paper: given a closed droplet contour, how should it be sampled to maximize the fidelity of the estimated area? In OLED inkjet printing, droplets are deposited into pixel wells with predefined geometry, and the projected droplet area directly reflects the filling state of the well. Insufficient projected area may correspond to incomplete filling, whereas excessive area increases the risk of overflow. Accurate area estimation from a limited number of contour samples is therefore of direct practical relevance to droplet observation in this process, providing the application basis for the area-driven sampling formulation developed in this study. Although droplet observation provides the primary application context, the formulation itself is defined on a sufficiently smooth fitted planar closed contour and does not depend on a droplet-specific shape model. Random closed curves are therefore included to evaluate the sampling law beyond the particular geometries observed in the droplet dataset. A limited contour-point budget is particularly relevant to high-frame-rate observation and resource-constrained contour representations, where the retained boundary points directly affect storage, transmission, and subsequent computational cost. Classical studies on contour discretization have investigated contour simplification, curve reconstruction, and adaptive sampling or representation. Contour-simplification methods reduce the number of contour points under geometric approximation criteria [8,9,10,11,12,13,14], curve-reconstruction methods recover continuous curves from sparse, unordered, or noisy samples [15,16,17,18], and adaptive sampling methods allocate samples or control points according to local geometric information to improve approximation fidelity and representation efficiency [19,20,21,22,23]. More recent studies have further extended adaptive contour representation through parametric, learning-based, and hierarchical formulations [24,25,26,27,28,29,30,31]. Among these directions, contour simplification and adaptive sampling are the most directly comparable to the present problem in terms of their input–output form: they take a contour representation as input and produce a finite set of representative points or a polygonal approximation. The main distinction lies in the optimization objective. Existing methods primarily allocate or retain points to control geometric quantities such as point-to-curve deviation, chord error, curvature representation, or overall reconstruction fidelity, which is appropriate when the sampled contour is intended to preserve geometric shape. In the droplet-observation problem considered here, the retained contour points are subsequently used to estimate the enclosed projected area. The present method therefore retains the same general contour-to-finite-point representation framework but derives the sampling distribution directly from the area error introduced by polygonal discretization. Nevertheless, these studies primarily optimize quantities such as reconstruction error, control-point efficiency, geometric smoothness, or visual fidelity. They do not directly address the error in the area estimation induced by contour discretization. This distinction is particularly important for droplet observation, as a minimum reconstruction error does not necessarily imply a minimum area error. For area estimation, contour-based droplet measurement has repeatedly emphasized the close relationship between extracted boundaries and projected-area or volume computation [4,5,6,7]. Even so, most approaches treat sampling as a mere implementation step following contour extraction rather than as an independent optimization objective. In other words, the literature has largely focused on improving the reconstruction quality of closed contours, but has far less frequently addressed the design of contour sampling mechanisms specifically aimed at more accurate area estimation. Area preservation has also been investigated in parametric and polygonal approximation [32,33,34,35]. These approaches preserve area through geometric operations such as parametric transformation, vertex relocation, edge movement, segment collapse, or the introduction of new vertices. In the present problem, however, the fitted contour itself is fixed and only the locations of a prescribed number of samples on that contour are allowed to vary. The geometric operations on which these area-preserving methods rely are therefore not admissible under the fixed-contour sampling constraint; directly restricting their output vertices to the fitted contour would change their original optimization mechanism rather than provide an equivalent sampling formulation. Motivated by these observations, this paper reformulates the processing of arbitrary closed droplet contours as an area-driven sampling problem. The objective is neither to produce a visually smoother curve, nor to achieve a more compact representation, nor to obtain a curve closer to the original contour; instead, under a fixed sampling budget, it aims to maximize the accuracy of area estimation from the contour sampling. The proposed method is formulated directly with respect to area estimation error. Starting from the leading-order local arc–chord area error, the sampling density is derived by minimizing the corresponding leading-order error functional; this optimality is interpreted asymptotically. The same global area-error model yields a leading-order estimate of the sampling budget associated with a prescribed area-error tolerance. The experiments compare the proposed method with representative sampling methods under identical sampling budgets and report area estimation accuracy together with geometric-fidelity metrics used to assess possible geometric degradation. 2. Related Work The related literature can be organized into five categories: contour simplification, curve reconstruction, adaptive sampling and contour representation, contour-based droplet measurement, and area-preserving approximation. (1)Contour simplification aims to reduce the number of contour points while preserving the geometric shape of the boundary as much as possible. The early work of [8,9] established the basic paradigm of recursively splitting the curve at the point of maximum perpendicular distance. Ref. [10] detected dominant points by combining curvature and cornerity measures with a region of support. Ref. [11] introduced a simplification criterion that iteratively eliminates points with the smallest contribution. Ref. [12] developed a dynamic algorithm that minimizes the maximum approximation error for a fixed number of segments. Ref. [13] extended optimal polygonal approximation to closed discrete curves by handling cyclic boundary conditions. Ref. [14] proposed a greedy approximation strategy that progressively removes redundant points while stabilizing the remaining vertices. Contour simplification aims to reduce the number of contour points while preserving the geometric shape of the boundary as much as possible. The early work of [8,9] established the basic paradigm of recursively splitting the curve at the point of maximum perpendicular distance. Ref. [10] detected dominant points by combining curvature and cornerity measures with a region of support. Ref. [11] introduced a simplification criterion that iteratively eliminates points with the smallest contribution. Ref. [12] developed a dynamic algorithm that minimizes the maximum approximation error for a fixed number of segments. Ref. [13] extended optimal polygonal approximation to closed discrete curves by handling cyclic boundary conditions. Ref. [14] proposed a greedy approximation strategy that progressively removes redundant points while stabilizing the remaining vertices. (2)Curve reconstruction aims to recover continuous and plausible curves from extreme conditions [15]. One representative work reconstructs curves from unorganized points by combining moving least squares with Euclidean minimum spanning tree refinement [16]. To handle noisy samples, another method exploits crust and medial axis filtering [17]. Sampling sufficiency for reliable reconstruction has also been systematically analyzed [18]. Curve reconstruction aims to recover continuous and plausible curves from extreme conditions [15]. One representative work reconstructs curves from unorganized points by combining moving least squares with Euclidean minimum spanning tree refinement [16]. To handle noisy samples, another method exploits crust and medial axis filtering [17]. Sampling sufficiency for reliable reconstruction has also been systematically analyzed [18]. (3)Adaptive sampling and contour representation seeks to improve geometric fidelity and representation efficiency under limited samples or control points. In adaptive sampling, ref. [19] pointed out the inefficiency of uniform sampling under high accuracy requirements, and ref. [20] further systematized curvature-based sampling for curves and surfaces. More recently, refs. [21,22] proposed near-optimal strategies based on deviation analysis and sectional error equivalence, respectively, while ref. [23] introduced an adaptive strategy based on maximum corresponding point deviation. Parametric and learned contour representations include persistent-homology-based reconstruction [24], learning-based control point prediction [25], neural-network-based knot solving for B-spline approximation [26], contour reconstruction based on concavity and convexity [27], perceptually faithful reconstruction by combining persistent homology with principal curves [28], image-processing-driven modeling and reconstruction via dual-channel detection and B-spline analysis [29], hierarchical local contour representation in AdaContour [30], and learning-based vector simplification with geometric-property evaluation [31]. Adaptive sampling and contour representation seeks to improve geometric fidelity and representation efficiency under limited samples or control points. In adaptive sampling, ref. [19] pointed out the inefficiency of uniform sampling under high accuracy requirements, and ref. [20] further systematized curvature-based sampling for curves and surfaces. More recently, refs. [21,22] proposed near-optimal strategies based on deviation analysis and sectional error equivalence, respectively, while ref. [23] introduced an adaptive strategy based on maximum corresponding point deviation. Parametric and learned contour representations include persistent-homology-based reconstruction [24], learning-based control point prediction [25], neural-network-based knot solving for B-spline approximation [26], contour reconstruction based on concavity and convexity [27], perceptually faithful reconstruction by combining persistent homology with principal curves [28], image-processing-driven modeling and reconstruction via dual-channel detection and B-spline analysis [29], hierarchical local contour representation in AdaContour [30], and learning-based vector simplification with geometric-property evaluation [31]. (4)Contour-based droplet measurement uses image-extracted droplet boundaries to obtain projected area, equivalent geometric descriptors, and volume-related quantities for droplet observation and process analysis [4,5,6,7]. Contour-based droplet measurement uses image-extracted droplet boundaries to obtain projected area, equivalent geometric descriptors, and volume-related quantities for droplet observation and process analysis [4,5,6,7]. (5)Area-preserving approximation has been investigated in parametric and polygonal representations. Area-preserving transformations have been considered in NURBS and CAD [32]. Bose et al. studied area-preserving approximations of polygonal paths [33], Buchin et al. developed area-preserving simplification and schematization through edge-move operations [34], and Kronenfeld et al. proposed an area-preserving segment-collapse strategy [35]. Area-preserving approximation has been investigated in parametric and polygonal representations. Area-preserving transformations have been considered in NURBS and CAD [32]. Bose et al. studied area-preserving approximations of polygonal paths [33], Buchin et al. developed area-preserving simplification and schematization through edge-move operations [34], and Kronenfeld et al. proposed an area-preserving segment-collapse strategy [35]. Overall, the most directly related contour-simplification and adaptive-sampling methods are primarily designed to preserve geometric fidelity or representation efficiency, whereas contour-based droplet observation ultimately uses the retained boundary samples to estimate projected area. This difference in downstream objective makes sampling-induced area error a more directly aligned criterion for the droplet-observation setting considered here, motivating the area-driven sampling formulation developed in the following section. 3. Methods 3.1. Contour Extraction and Closed-Curve Representation The proposed method takes an ordered closed contour obtained after image processing as input. In the droplet dataset experiments, image acquisition, preprocessing, binary segmentation, and dense contour extraction precede the sampling stage, which starts from the resulting ordered contour. After duplicate points are removed and the contour orientation is normalized, the contour is represented by a periodic cubic spline. For a parametric contour,(1)r(t)=(x(t),y(t)),t∈[0,T], subject to(2)r(0)=r(T). The ordered contour is represented by a periodic cubic spline, resulting in a C2 closed curve with well-defined tangent, arc length, and curvature [36,37]. This smooth representation provides the geometric quantities required by the subsequent area-error analysis and sampling-density construction. Since curvature is determined by first- and second-order derivatives of the fitted curve, its numerical estimate naturally depends on the contour representation and spline fitting. The same fitted spline is used as the common contour representation for all sampling methods in the experiments. 3.2. Problem Formulation for Droplet Area Estimation Let Ω⊂R2 denote the region enclosed by the fitted closed contour, and let its counterclockwise boundary be parameterized by arc length as(3)r(s)=(x(s),y(s)),s∈[0,L], where L is the total arc length. Its enclosed reference area is denoted by(4)A=∫∫ΩdA. For a prescribed sampling budget of N distinct points on a closed curve, the sampling parameters are defined by(5)0≤s1