Fact-checked by Grok 2 weeks ago

Image gradient

In image processing and , the image gradient is defined as the of the values between neighboring s, quantifying the rate and of intensity or color change across an image at each point. It is typically represented as a two-dimensional consisting of (\partial I / \partial x) and vertical (\partial I / \partial y) partial derivatives, where the \|\nabla I\| = \sqrt{\left( \frac{\partial I}{\partial x} \right)^2 + \left( \frac{\partial I}{\partial y} \right)^2} measures strength and the \theta = \atan2\left( \frac{\partial I}{\partial y}, \frac{\partial I}{\partial x} \right) indicates the of the most rapid intensity change, often perpendicular to s. These derivatives are approximated using discrete with kernels, such as filters, to handle the continuous nature of images on digital grids. The image gradient serves as a foundational tool for detecting discontinuities in , enabling the identification of , boundaries, and textures that delineate objects within scenes. Its computation is essential for illumination-invariant features, as gradient distributions remain robust to uniform lighting changes, making it valuable in applications like face recognition and . Beyond basic , gradients underpin advanced techniques such as the (HOG) descriptor, which encodes local gradient orientations for in models. Common methods for estimating image gradients include the , introduced in 1968, which applies 3×3 convolution masks to compute smoothed approximations of the derivatives, balancing with edge localization. Variants like Prewitt and Scharr operators offer similar functionality but differ in weights for or sensitivity, while Gaussian derivatives provide scale-invariant gradients by convolving with a Gaussian before . These approaches highlight the gradient's versatility, from in hardware implementations to integration in pipelines for feature extraction.

Mathematical Foundations

Continuous Case

In the continuous case, an image is modeled as a two-dimensional scalar f(x, y) that assigns an value to each point in the continuous . The of this , denoted \nabla f(x, y), is a defined as \nabla f(x, y) = \begin{pmatrix} \frac{\partial f}{\partial x} \\ \frac{\partial f}{\partial y} \end{pmatrix}, where \frac{\partial f}{\partial x} and \frac{\partial f}{\partial y} are the partial derivatives representing the instantaneous rates of change of along the horizontal and vertical directions, respectively. These partial derivatives presuppose basic knowledge of , capturing how varies independently with respect to each coordinate. The magnitude of the gradient, |\nabla f|, quantifies the overall rate of intensity change and is derived from the norm of the gradient vector: |\nabla f| = \sqrt{ \left( \frac{\partial f}{\partial x} \right)^2 + \left( \frac{\partial f}{\partial y} \right)^2 }. The orientation \theta indicates the direction of this change and is given by \theta = \atantwo\left( \frac{\partial f}{\partial y}, \frac{\partial f}{\partial x} \right), where \atantwo is the two-argument arctangent that accounts for the correct . From a perspective, the \nabla f points in the direction of the steepest ascent of the at each point, with its magnitude specifying the slope of that ascent. This aligns with the concept of directional derivatives, where the rate of change in an arbitrary direction \mathbf{u} = (u_x, u_y) (a ) is the \nabla f \cdot \mathbf{u} = \frac{\partial f}{\partial x} u_x + \frac{\partial f}{\partial y} u_y, achieving its maximum value along the direction itself. To illustrate, consider a simple linear f(x, y) = ax + by + c, where a and b determine the components. The partial derivatives are constants \frac{\partial f}{\partial x} = a and \frac{\partial f}{\partial y} = b, yielding a uniform (a, b) everywhere, which reflects a steady, directionally consistent increase. For a more complex example, a Gaussian centered at the origin, f(x, y) = e^{-\frac{x^2 + y^2}{2\sigma^2}} (with \sigma as the standard deviation), has \nabla f(x, y) = -\frac{(x, y)}{\sigma^2} f(x, y). Here, the points radially inward toward the center of highest , with magnitude peaking at intermediate distances before diminishing, demonstrating how the encodes localized variations around a smooth peak.

Discrete Case

In digital image processing, images are represented as two-dimensional arrays f[i,j], where i and j denote discrete pixel coordinates, and each entry holds a finite intensity value from a quantized range. The discrete gradient at each pixel is defined as a vector \mathbf{g}[i,j] = [g_x[i,j], g_y[i,j]], providing approximations of the partial derivatives with respect to the horizontal and vertical directions, respectively. This vector representation captures the local rate and direction of intensity change, serving as the discrete analog to the continuous gradient operator. To approximate these partial derivatives on the pixel grid, finite difference methods are employed, assuming uniform spacing between pixels. The forward difference for the x-component is given by g_x[i,j] \approx \frac{f[i+1,j] - f[i,j]}{\Delta x}, where \Delta x is the pixel spacing in the x-direction, often normalized to 1 for simplicity. The backward difference uses g_x[i,j] \approx \frac{f[i,j] - f[i-1,j]}{\Delta x}, while the central difference, which offers better symmetry and reduced bias, is g_x[i,j] \approx \frac{f[i+1,j] - f[i-1,j]}{2 \Delta x}. Analogous formulas apply to the y-component by swapping indices. Normalization by \Delta x and \Delta y ensures scale invariance when pixel spacing varies, such as in images with anisotropic resolution. Discretization introduces artifacts, notably , where of high-frequency components causes spatial misalignment in estimates, such as half-pixel shifts in kernels. This can lead to direction-dependent errors in the and . To mitigate such issues, isotropic approximations are necessary, ensuring that the response remains consistent across rotations and avoids preferential toward axis-aligned directions. For color images in RGB format, gradients are typically computed channel-wise, yielding separate vectors for the red, green, and blue components, as the channels represent independent intensity signals. These can then be combined, for instance, by averaging or using a luminance-weighted formula, to obtain a perceptual gradient. Mathematically, the discrete gradient functions as a linear operator on the image array, satisfying superposition: the gradient of a linear combination of images equals the corresponding combination of their gradients. This property aligns it with differentiation in two-dimensional discrete signal processing, where images are treated as sampled signals and the gradient extracts edge-related frequency components.

Computation Methods

Finite Difference Approximations

Finite difference approximations provide a straightforward numerical method to compute the image gradient on discrete pixel grids by estimating partial derivatives through differences in intensity values. These methods are foundational for gradient computation in digital images, where the image function f(i,j) represents intensity at pixel coordinates (i,j). The horizontal gradient component \frac{\partial f}{\partial i} and vertical component \frac{\partial f}{\partial j} are approximated using nearby pixel values, assuming uniform pixel spacing \Delta i = \Delta j = 1. First-order finite difference schemes offer simple implementations but vary in accuracy. The forward difference for the horizontal gradient at pixel (i,j) is derived from the definition of the as \frac{\partial f}{\partial i} \approx f(i+1,j) - f(i,j), which follows from the first-order f(i+1,j) = f(i,j) + \frac{\partial f}{\partial i} \cdot 1 + O(1^2), yielding a of O(1). Similarly, the backward difference uses \frac{\partial f}{\partial i} \approx f(i,j) - f(i-1,j), with the same O(1) error from the f(i-1,j) = f(i,j) - \frac{\partial f}{\partial i} \cdot 1 + O(1^2). For the vertical direction, analogous approximations apply: forward \frac{\partial f}{\partial j} \approx f(i,j+1) - f(i,j) and backward \frac{\partial f}{\partial j} \approx f(i,j) - f(i,j-1). These schemes are computationally efficient but sensitive to noise due to their first-order accuracy. To improve accuracy, central difference schemes employ symmetric approximations, reducing the truncation error. The central difference for the horizontal gradient is \frac{\partial f}{\partial i} \approx \frac{f(i+1,j) - f(i-1,j)}{2}, derived by subtracting Taylor expansions: f(i+1,j) = f(i,j) + \frac{\partial f}{\partial i} + \frac{1}{2} \frac{\partial^2 f}{\partial i^2} + O(1^3) and f(i-1,j) = f(i,j) - \frac{\partial f}{\partial i} + \frac{1}{2} \frac{\partial^2 f}{\partial i^2} + O(1^3), so their difference divided by 2 cancels the second-order term, bounding the error at O(1^2). The vertical central difference is \frac{\partial f}{\partial j} \approx \frac{f(i,j+1) - f(i,j-1)}{2}, with equivalent error analysis. These second-order methods provide better estimates for smooth intensity profiles but require pixels on both sides, complicating boundary handling. Boundary conditions are essential for pixels near image edges where neighboring values may lie outside the domain. Common approaches include zero-padding, which sets external values to zero (equivalent to Dirichlet boundary conditions); replication, which extends edge pixels constantly outward; or mirroring, which reflects the image across the (akin to Neumann conditions for zero normal ). These methods ensure gradient computation across the entire image but can introduce artifacts, such as artificial edges from zero-padding. The of computation is O(N) for an with N , as each component requires a constant-time difference per pixel. Basic for computing the (\frac{\partial f}{\partial i}, \frac{\partial f}{\partial j}) using central differences, assuming zero-padding for boundaries, is as follows:
for i from 1 to M-2:  # M rows
    for j from 1 to N-2:  # N columns
        Gx[i,j] = (f[i+1,j] - f[i-1,j]) / 2
        Gy[i,j] = (f[i,j+1] - f[i,j-1]) / 2
# Handle boundaries separately with forward/backward or padding
For edge pixels, forward or backward differences apply with the chosen extension. As an illustrative example, consider a 1D profile f(k) = \sin(2\pi k / 10) for k = 0 to $9. The true at k=1 is \cos(2\pi \cdot 1 / 10) \cdot (2\pi / 10) \approx 0.51. A forward difference yields (f(2) - f(1))/1 \approx 0.36 (error ~29%), backward (f(1) - f(0))/1 \approx 0.59 (error ~16%), and central (f(2) - f(0))/2 \approx 0.48 (error ~6%), demonstrating the improved accuracy of the central scheme.

Convolution-Based Operators

Convolution-based operators estimate the image gradient by convolving the input with small designed to approximate partial derivatives while incorporating to reduce sensitivity. In this framework, the horizontal component g_x is computed as the of the f with a K_x, denoted g_x = f * K_x, and similarly for the vertical component g_y = f * K_y. The overall is then |\nabla f| = \sqrt{g_x^2 + g_y^2}, providing an strength measure. These typically span 2×2 or 3×3 neighborhoods, extending approximations by averaging over multiple pixels for enhanced robustness. The Prewitt , introduced in , uses weights in its kernels to detect and vertical edges. For the x-direction, the kernel is K_x = \frac{1}{3} \begin{bmatrix} -1 & 0 & 1 \\ -1 & 0 & 1 \\ -1 & 0 & 1 \end{bmatrix}, which averages differences across rows, effectively combining with . The corresponding y-kernel is transposed. This design promotes for cardinal directions but treats all neighboring pixels equally, leading to moderate noise suppression. The , developed in , refines this approach with central weighting to better approximate Gaussian smoothing, improving noise resistance and edge localization. Its x-kernel is K_x = \begin{bmatrix} -1 & 0 & 1 \\ -2 & 0 & 2 \\ -1 & 0 & 1 \end{bmatrix}, where the factor of 1/8 is often applied post-convolution for normalization, though omitted here for integer arithmetic. The y-kernel is similarly transposed. This weighting emphasizes the central row, yielding a smoother that attenuates high-frequency noise more effectively than the Prewitt while maintaining good . In the , the Sobel's exhibits less shift for edge orientations compared to simpler operators, reducing localization errors. For diagonal edges, the Roberts cross operator from employs compact 2×2 , prioritizing computational speed over smoothing. The primary kernel for one diagonal is K_1 = \begin{bmatrix} 1 & 0 \\ 0 & -1 \end{bmatrix}, with the orthogonal kernel K_2 = \begin{bmatrix} 0 & 1 \\ -1 & 0 \end{bmatrix} for the other direction. These detect 45-degree edges but are highly sensitive to noise due to minimal averaging and anisotropic response, performing poorly on horizontal/vertical features. In , Roberts kernels show pronounced phase shifts and weaker low-frequency suppression, amplifying noise in practical images. Comparisons reveal trade-offs: Sobel offers superior noise immunity and near-isotropic response across orientations, ideal for general use; Prewitt balances simplicity and performance but with slightly higher noise sensitivity; Roberts excels in speed for low-noise scenarios but suffers from anisotropy and amplification of artifacts. Frequency-domain evaluation confirms Sobel's smoother magnitude response and reduced phase distortion, enhancing edge accuracy over Prewitt's uniform filter and Roberts' sharp but directional cutoff. Implementation efficiency is boosted by separability in operators like Sobel, where the 2D convolution decomposes into sequential 1D horizontal and vertical passes (e.g., row-wise with [ -1, 0, 1 ] followed by column-wise with [ 1, 2, 1 ]^T, up to scaling), reducing complexity from O(9N) to O(4N) per component for an N-pixel image. For color images, gradients are often computed on (e.g., weighted sum of RGB channels) to capture perceptual edges, though gradients across channels can detect chromatic boundaries.

Applications

Edge Detection

In image processing, edge detection identifies boundaries between regions of differing intensity, and image gradients are fundamental to this process as they quantify the rate of intensity change. The magnitude of the gradient serves as an indicator of edge strength, with higher values signaling abrupt transitions typical of edges, while the orientation of the gradient vector reveals the direction perpendicular to the edge. This dual information enables algorithms to not only locate potential edges but also align them accurately, distinguishing true boundaries from gradual variations or noise. The historical development of gradient-based edge detection began in the 1960s with Lawrence Roberts' introduction of the Roberts cross operator in 1963, which used simple 2x2 kernels to approximate the first-order gradient along diagonal directions for basic edge enhancement. In the late 1960s and 1970s, the Sobel operator, proposed by Irwin Sobel and Gary M. Feldman in 1968, advanced this by employing 3x3 isotropic kernels that combined gradient approximation with smoothing to reduce noise sensitivity, making it a widely adopted method for computing horizontal and vertical derivatives. These early first-derivative approaches laid the groundwork for more sophisticated techniques, culminating in John Canny's 1986 algorithm, which optimized edge detection criteria for low error rates, good localization, and single-response to edges. The Sobel operator remains a common gradient estimator in these pipelines due to its balance of simplicity and effectiveness. The exemplifies the use of in a structured to produce thin, continuous edges. It starts with Gaussian to reduce , followed by computation—typically via the —to obtain the magnitude and direction at each . Non-maximum suppression then refines the edge map by interpolating along the direction and suppressing that are not local maxima, ensuring edges are one pixel thick. thresholding follows, applying a high (e.g., 150 for 8-bit images) to identify strong edges and a low (e.g., 50) for weak ones, retaining weak edges only if they connect to strong edges via 8-connectivity, thus minimizing breaks in contours. In contrast, the Marr-Hildreth approach from 1980 employs second derivatives for edge detection, convolving the image with a Laplacian of Gaussian (LoG) filter to detect zero-crossings, which mark inflection points in intensity profiles. This method ties briefly to first-derivative gradients, as LoG zero-crossings often align with maxima in gradient magnitude, providing a multi-scale perspective on edges robust to fine details. Evaluation of gradient-based edge detectors relies on metrics like precision, which quantifies the proportion of detected edges that match ground-truth boundaries (true positives over true positives plus false positives), and recall, which measures the proportion of actual edges captured (true positives over true positives plus false negatives). These metrics highlight challenges such as noise-induced false edges, particularly in first-derivative methods like Roberts and Sobel, where Gaussian noise can produce spurious gradient peaks, lowering precision in low-contrast or textured regions. As an illustrative example, consider a synthetic image with a vertical step separating a dark region ( 0) from a bright one ( 255), such as a simple boundary. First, compute the using the : the horizontal yields near-zero values away from the but a sharp peak ( ≈255) at the transition pixels, while the vertical remains low. The map thus highlights a vertical line of high values at the location, with orientation consistently horizontal. Applying non-maximum suppression thins this to a single-pixel-wide line by discarding adjacent lower- pixels. Finally, hysteresis thresholding with low (50) and high (150) thresholds confirms the continuous , suppressing any minor fluctuations while preserving the full boundary. This step-by-step process demonstrates how gradients isolate and refine s in controlled scenarios, forming the basis for real-image applications.

Feature Extraction and Matching

Image gradients play a crucial role in feature extraction by quantifying local intensity changes, which highlight edges, corners, and textures essential for identifying distinctive points in images. These features facilitate matching between images taken from varying viewpoints, scales, or under different lighting conditions, supporting tasks like in cluttered scenes and stereo matching for . By encoding gradient magnitudes and orientations, methods transform raw pixel data into compact, invariant representations that resist geometric and photometric distortions. The (HOG) descriptor captures the distribution of directions within local image patches to represent object shapes robustly. It computes orientations and magnitudes at each , then bins these orientations into histograms over small spatial s, typically aggregating 8-9 bins per for coarse . Block-level follows to mitigate illumination variations, ensuring invariance across the image. Introduced by Dalal and Triggs in 2005 for detection, HOG excels in scenarios requiring dense feature coverage, such as human pose estimation, due to its simplicity and effectiveness in -based shape encoding. The (SIFT) leverages image for both keypoint detection and description, achieving invariance to scale and . Keypoints are identified as extrema in the () scale space, which approximates the Laplacian of Gaussian for , followed by precise localization using information to sub-pixel accuracy. A dominant is assigned by analyzing magnitudes in a neighborhood around each keypoint, enabling . The descriptor is then formed by computing a 128-dimensional from histograms of orientations and magnitudes in a 16x16 of sub-regions, weighted by a Gaussian envelope for locality. Developed by Lowe in , SIFT has become a for feature matching in and panoramic stitching, balancing distinctiveness with robustness to affine transformations. Speeded-Up Robust Features (SURF) approximates SIFT's gradient-based operations for accelerated computation while preserving and invariance. It detects interest points using the of the , approximated via box filters computed efficiently with integral images, which reduce costs. Orientation assignment employs responses in horizontal and vertical directions to determine a reproducible dominant angle from gradient-like responses. The descriptor aggregates these wavelet coefficients into a 64-dimensional vector over a 4x4 sub-region grid, with sub-sampling for invariance. Proposed by Bay et al. in 2006, SURF offers up to three times faster performance than SIFT on standard benchmarks like image matching, making it suitable for real-time applications such as . Feature matching relies on comparing these gradient-derived descriptors to establish correspondences between images. Common techniques include nearest-neighbor search using for SIFT and descriptors, often refined with a to discard ambiguous matches by ensuring the closest neighbor's distance is significantly smaller than the second-closest. For , correlation-based matching aligns histogram vectors across patches. These methods enable accurate point-to-point associations, crucial for tasks like estimation in wide-baseline matching. In motion tracking, gradients underpin estimation by assuming brightness constancy and spatial smoothness, allowing displacement computation from intensity derivatives. The Lucas-Kanade method solves for flow vectors in a least-squares framework over small windows, using the image tensor to constrain solutions under the small-motion assumption. Introduced by Lucas and Kanade in , this local approach integrates information for sub-pixel flow fields, widely applied in video stabilization and structure-from-motion pipelines. These gradient-based techniques provide strong invariance to scale and rotation—SIFT and repeatably match features across 2-4 octave scales and 360-degree rotations with over 90% accuracy on benchmark datasets—while offers partial illumination robustness via normalization. Computational trade-offs are evident: requires operations for dense grids, suitable for offline processing, whereas 's integral image approximations yield speeds on standard , though at minor cost to distinctiveness compared to SIFT's exhaustive sampling.

Segmentation and Advanced Uses

Image gradients play a crucial role in segmentation techniques that partition images into meaningful regions by leveraging gradient magnitude to delineate boundaries. In active contour models, commonly known as , deformable curves evolve under the influence of gradient flow to align with high-gradient edges in the image. These models minimize an energy functional that balances internal smoothness constraints with external forces derived from the image gradient, where the external term incorporates the magnitude of the gradient, |\nabla I|, to attract the contour toward object boundaries. Introduced by Kass, Witkin, and Terzopoulos in their seminal work, this approach enables interactive segmentation by allowing user constraints to guide the contour near features of interest. Watershed segmentation treats the magnitude of an as a topographic surface, simulating flooding from minima to partition the into catchment basins that correspond to regions separated by ridges of high . This method, originally proposed by Beucher and Lantuejoul, effectively merges homogeneous regions but often results in over-segmentation due to noise-induced minima. To mitigate this, marker-controlled watershed variants impose internal and external markers to guide the flooding process, restricting basins to predefined seeds and reducing fragmentation while preserving -based boundaries. and Soille's immersion further refined this by processing the landscape in increasing order of height, enabling efficient computation for large images. In modern frameworks, image inform convolutional neural networks (CNNs) through feature maps in early layers, which detect edge-like patterns akin to traditional operators, serving as foundational inputs for higher-level segmentation tasks. Post-2010s developments integrate into training via , where -based losses optimize network parameters for precise boundary delineation in semantic segmentation. For instance, the architecture employs a -descent-based optimization with loss to train encoder-decoder networks on biomedical images, achieving pixel-wise segmentation by propagating information through skip connections that preserve spatial details from -sensitive maps. This addresses gaps in traditional methods by enabling end-to-end learning from -derived inputs, as demonstrated in Ronneberger et al.'s work on volumetric data. Beyond , image gradients find applications in geographic information systems (GIS) for analyzing and demographic flows. In digital elevation models (DEMs), gradients compute and from data, representing the rate of change in height to model steepness and drainage patterns essential for hydrological simulations. For visualization, gradient-based flow maps illustrate or density gradients, using fields derived from spatial gradients to depict directional changes in distribution over raster grids. Such techniques, as explored in spatiotemporal mobility studies, enhance by highlighting gradient-driven flows in densely textured regions. Despite these advances, challenges persist in handling weak gradients within textured regions, where subtle intensity variations lead to ambiguous boundaries and incomplete segmentations in methods like or . Computational demands for gradient computations in high-resolution images are addressed through GPU acceleration, which parallelizes or operations to speed up active contour evolution and watershed flooding by orders of magnitude. Reviews of GPU-based segmentation highlight up to 100-fold speedups for level-set methods incorporating gradient vector flows, making real-time applications feasible in . Gradient-derived features, such as those from descriptors, can briefly augment inputs to these models for improved robustness in textured scenes.

References

  1. [1]
    Image Gradient - an overview | ScienceDirect Topics
    Image gradient refers to the derivative of the pixel values of neighboring pixels in an image. It is used as a measure of image sharpness, where sharper ...
  2. [2]
    [PDF] 4.0 Image Gradients and Gradient Filtering
    How do you compute the image gradient? Choose a derivative filter. Run filter over image. Image gradient. What are the dimensions? What ...
  3. [3]
    [PDF] Sobel Operator and Canny Edge Detector ECE 480 Fall 2013 Team ...
    Out of numerous edge detection methods, this paper discussed and analyzed the strengths and weaknesses of Sobel and Canny edge detection methods. Sobel operator ...
  4. [4]
    [PDF] Hardware Implementation of Sobel Edge Detection Algorithm
    In this work we are using FPGA boards which have many advantages like faster time to plug , par- allel task performance, adaptibility, real time applications.
  5. [5]
    [PDF] Lecture 2: Image filtering - CS@Cornell
    Image gradient. • The gradient of an image: The edge strength is given by the gradient magnitude: The gradient direction is given by: • how does this relate ...
  6. [6]
    18 Image Derivatives - Foundations of Computer Vision
    When processing images we have to use discrete approximations for the Gaussian derivatives. After discretization, many of the properties of the continuous ...18 Image Derivatives · 18.5 Gaussian Derivatives · 18.9 Image Laplacian
  7. [7]
    [PDF] Digital Image Processing
    This edition is the most comprehensive revision of Digital Image Processing since the book first appeared in 1977.As the 1977 and 1987 editions by Gonzalez and ...
  8. [8]
    [PDF] A Review of Classic Edge Detectors - IPOL Journal
    Jun 16, 2015 · In the 70's, edge detection methods were implemented using small operators (such as Sobel masks), attempting to compute an approximation of the ...
  9. [9]
    [PDF] Object enhancement and extraction - UTK-EECS
    Judith M. S. Prewitt. University of Pennsylvania. Philadelphia, Pennsylvania. INTRODUCTION. Object enhancement, extraction, characterization and recognition are.
  10. [10]
    (PDF) A 3×3 isotropic gradient operator for image processing
    A 3x3 isotropic gradient operator for image processing, presented at the Stanford Artificial Intelligence Project (SAIL) in 1968.
  11. [11]
    Beginner's Guide To Image Gradient - Analytics Vidhya
    Nov 12, 2024 · A gradient in an image refers to the change in intensity or colour between neighbouring pixels. It represents how quickly pixel values are ...What is an Image Gradient? · Mathematical Calculation of... · Overview of Filters
  12. [12]
    (PDF) An Isotropic 3x3 Image Gradient Operator - ResearchGate
    PDF | On Feb 8, 2014, Irwin Sobel published An Isotropic 3x3 Image Gradient Operator | Find, read and cite all the research you need on ResearchGate.
  13. [13]
  14. [14]
  15. [15]
    Edge Detection Process - an overview | ScienceDirect Topics
    Quantitative evaluation of edge detection relies on objective and reproducible metrics, with precision, recall, F1-score, average precision (AP), and P measure ...
  16. [16]
    [PDF] Chapter 5. Edge Detection
    The figures at the end of this section show the perfor- mance of this operator. The Sobel operator is one of the most commonly used edge detectors. -1 0 1. -2 0 ...
  17. [17]
    [PDF] Histograms of Oriented Gradients for Human Detection
    After reviewing existing edge and gra- dient based descriptors, we show experimentally that grids of Histograms of Oriented Gradient (HOG) descriptors sig-.
  18. [18]
    [PDF] Distinctive Image Features from Scale-Invariant Keypoints
    Jan 5, 2004 · This paper presents a method for extracting distinctive invariant features from images that can be used to perform reliable matching between ...
  19. [19]
    [PDF] SURF: Speeded Up Robust Features
    In this paper, we present a novel scale- and rotation-invariant interest point detector and descriptor, coined SURF (Speeded Up Ro- bust Features). It ...
  20. [20]
    [PDF] An Iterative Image Registration Technique - CMU Robotics Institute
    Image registration finds a variety of applications in computer vision, such as image matching for stereo vision, pattern recognition, and motion analysis.
  21. [21]
    Snakes: Active contour models | International Journal of Computer ...
    We have used snakes successfully for interactive interpretation, in which user-imposed constraint forces guide the snake near features of interest. Article PDF ...
  22. [22]
    U-Net: Convolutional Networks for Biomedical Image Segmentation
    May 18, 2015 · In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more ...Missing: gradient- loss
  23. [23]
    A systematic review and meta-analysis of Digital elevation model ...
    Besides the values of height, a DTM consists of other topographic attributes or terrain morphological elements such as slope, aspect, curvature, gradient, ...
  24. [24]
    Visualizing population mobility from spatiotemporally aggregated ...
    Dec 11, 2022 · In this study, based on image-processing techniques, a two-step 3D gradient method is adopted to extract the movement features. The first step ...
  25. [25]
    Techniques and Challenges of Image Segmentation: A Review - MDPI
    Mar 2, 2023 · There are two difficulties in image segmentation: (1) how to define “meaningful regions”, as the uncertainty of visual perception and the ...
  26. [26]
    Medical image segmentation on GPUs – A comprehensive review
    This review investigates the use of GPUs to accelerate medical image segmentation methods. A set of criteria for efficient use of GPUs are defined.