Fact-checked by Grok 2 weeks ago

Line detection

Line detection is a fundamental task in that involves identifying and extracting straight line segments from digital images, representing linear structures such as edges of objects, boundaries, or geometric features in scenes. These line segments are typically parameterized by their endpoints, midpoint, length, and orientation angle, providing compact and descriptive representations of image content. The technique is essential for higher-level vision applications, including (), , , scene parsing, and pose estimation, as line segments offer structural and geometric cues that enable robust scene understanding and in and systems. Classical methods, such as the —originally patented in 1962 by Paul V. C. Hough for recognizing complex patterns in bubble chamber images—accumulate votes from edge points in a parameter space to detect lines, while later advancements like the Line Segment Detector (LSD), introduced in 2010 by Grompone von Gioi et al., achieve real-time performance with subpixel accuracy and no parameter tuning by grouping aligned edge pixels locally. In recent years, has revolutionized line detection, with methods like convolutional neural networks (CNNs) and hybrid approaches combining learned features with traditional techniques to improve accuracy, robustness to noise, and handling of complex . Notable examples include DeepLSD (2023), which uses a deep network to generate line attraction fields for precise segment refinement, and other learning-based detectors such as LCNN and MLSD, which leverage end-to-end training on datasets like Wireframe for wireframe and . Advancements have continued into 2024 and 2025, with transformer-based methods like RANK-LETR improving detection through learnable geometric priors. These developments have expanded applications to areas like document , indoor , and autonomous , where reliable line supports tasks from to environmental modeling.

Overview

Definition and Scope

Line detection is a core task in image processing and , focused on identifying straight lines—either infinite or finite segments—within digital images to extract structural features for higher-level analysis. These lines represent alignments of edge pixels that delineate object boundaries or scene geometries, often derived from edge maps produced by preliminary filtering. As a low-level feature extraction process, it provides essential geometric cues that support tasks like and scene reconstruction, emphasizing straightness and over curved or irregular forms. A key distinction exists between line detection, which typically targets infinite lines defined parametrically without endpoints, and line segment detection, which localizes finite, bounded segments with explicit start and end points to capture localized structures. Infinite line detection suits global scene interpretations, such as horizon estimation, while segment detection addresses practical needs like wireframe parsing in urban environments. This differentiation highlights the scope's emphasis on both parametric ideals and realistic, truncated representations in images. The field emerged in the early 1960s with the , a foundational method for detecting lines in images, amid efforts in to enable automated scene understanding, building on early techniques for processing pictorial data. Pioneering systems from this era, including extensions of edge-finding algorithms, laid the groundwork by addressing how machines could interpret linear primitives in complex visuals. Line detection faces inherent challenges, including high sensitivity to that disrupts collinear alignments, partial occlusions that fragment potential lines, and distortions that warp linear features across viewpoints. These issues complicate robust in real-world , where environmental variations often degrade accuracy and completeness.

Applications in

Line detection plays a pivotal role in autonomous driving systems, particularly for detection, where it identifies boundaries to ensure safe navigation and trajectory planning. In these applications, algorithms process camera feeds to extract markings, enabling vehicles to maintain position within lanes even under varying or conditions. For instance, learning-based detection networks have been integrated into advanced driver assistance systems (ADAS), improving vehicle control in scenarios. In document analysis, line detection facilitates skew correction by identifying text lines or page borders in scanned images, allowing for accurate alignment and preprocessing before (OCR). This process straightens distorted documents, enhancing readability and reducing errors in digital archiving or form processing. Techniques based on and profiles have proven effective for estimating angles in text-heavy pages. Architectural drawing interpretation relies on line detection to parse floor plans, elevations, and blueprints, converting line segments into semantic representations of structures like walls and doors. This aids in automated (BIM) and design validation, where lines represent edges of architectural elements. Semantic interpretation methods extract and classify these lines to reconstruct spatial relationships. In , line detection supports environment mapping by delineating structural features from sensor data, such as scans, to build grids or topological maps for navigation. Mobile robots use line segments to represent corridors or obstacles, enabling (SLAM) in indoor settings. Point-line fusion approaches enhance map accuracy in multi-robot collaborations. Beyond these primary uses, line detection enables estimation, which infers 3D scene geometry from 2D images by converging parallel lines toward focal points, supporting monocular in urban environments. Similarly, in aerial imagery, horizon line detection identifies the boundary between and ground, aiding attitude estimation and stabilization for unmanned aerial vehicles (UAVs). Block-based methods robustly detect these vanishing lines even in cluttered scenes. The evolution of line detection applications traces back to the early 1980s in , where classical methods like the enabled initial environment mapping and obstacle avoidance in mobile robots. By the 2020s, integration with has advanced real-time line overlay in (AR) and (VR), such as visualizing lane maps on windshields for enhanced driver awareness. Quantitative studies demonstrate the impact of line-based features on related tasks; for example, combining (HOG), which captures line orientations, with (LBP) in systems has achieved classification accuracies exceeding 93.6%, outperforming standalone edge detectors by leveraging structural line cues for shape discrimination.

Mathematical Foundations

Parametric Line Representations

In , lines in 2D images are commonly represented using parametric equations that express the relationship between image coordinates x and y. One standard form is the slope-intercept equation, y = mx + c, where m is the and c is the . This representation is intuitive for lines with finite slopes but fails for vertical lines, where m approaches , leading to numerical instability in parameter estimation. A more general parametric form is the Cartesian line equation ax + by + c = 0, where a, b, and c are constants, with a and b not both zero. To ensure uniqueness and facilitate computations such as distance measurements, this equation is often normalized such that a^2 + b^2 = 1, making the vector (a, b) a unit normal to the line and |c| the perpendicular distance from the origin to the line. The normal (or polar) form provides another parameterization, x \cos \theta + y \sin \theta = \rho, where \rho \geq 0 is the perpendicular distance from the origin to the line, and \theta \in [0, \pi) is the angle of the normal vector from the x-axis. This form avoids the issues of unbounded parameters in slope-intercept representations and is particularly suited for quantization in discrete parameter spaces. To derive the normal parameters from Cartesian coordinates, start with two points (x_1, y_1) and (x_2, y_2) on the line, yielding the general equation (y_1 - y_2)x + (x_2 - x_1)y + (x_1 y_2 - x_2 y_1) = 0. Normalizing so a^2 + b^2 = 1 gives a = \cos \theta and b = \sin \theta, with \rho = -c (adjusted for sign). For a point (x_i, y_i) on the line, the relation \rho = x_i \cos \theta + y_i \sin \theta traces a sinusoid in the \theta-\rho parameter space. While these representations model infinite lines, real image features are finite line segments, which extend the form by adding start and end points, such as (x_s, y_s) and (x_e, y_e), along with the supporting line equation to capture bounded extent and avoid ambiguities in length or position. These forms are typically applied to candidate points extracted from processes, such as gradient-based operators that highlight potential line locations.

Prerequisites: Edge Detection

Edge detection serves as a fundamental preprocessing step in line detection, identifying pixels that likely belong to lines by highlighting boundaries in an image. Edges are defined as significant local changes in image intensity, typically occurring at the boundaries between regions with different properties, such as object silhouettes or texture transitions. Common first-order derivative operators approximate the image gradient to detect these intensity discontinuities, with the Sobel, Prewitt, and Roberts operators being widely used due to their simplicity and efficiency. The Sobel operator employs 3×3 convolution kernels to estimate gradients in the horizontal and vertical directions; for example, the horizontal gradient kernel G_x is given by: G_x = \begin{bmatrix} -1 & 0 & 1 \\ -2 & 0 & 2 \\ -1 & 0 & 1 \end{bmatrix} while the vertical gradient kernel G_y is its transpose. The Prewitt operator uses similar 3×3 kernels but with uniform weights of 1 instead of the Sobel's central emphasis, such as: G_x = \begin{bmatrix} -1 & 0 & 1 \\ -1 & 0 & 1 \\ -1 & 0 & 1 \end{bmatrix} for horizontal edges, making it slightly less smoothing than Sobel. The Roberts operator, an earlier 2×2 kernel-based method, focuses on diagonal gradients for faster computation on simple hardware, using kernels like: G_x = \begin{bmatrix} 1 & 0 \\ 0 & -1 \end{bmatrix}, \quad G_y = \begin{bmatrix} 0 & 1 \\ -1 & 0 \end{bmatrix} though it is more prone to noise due to its smaller support. For more robust edge extraction, the Canny edge detector applies Gaussian smoothing to reduce noise, followed by gradient computation (often using Sobel kernels), non-maximum suppression to thin edges by retaining only local maxima in the gradient direction, and hysteresis thresholding with two thresholds to connect weak edges to strong ones, producing continuous, thin edge contours. The resulting output is a binary edge map, where edge pixels are marked (typically as 1 or white) and non-edge pixels are suppressed (0 or black), serving directly as input for subsequent line detection algorithms that accumulate or fit lines to these candidate points. Despite these advances, edge detection methods exhibit limitations that can propagate to line detection, including high sensitivity to image noise, which introduces false edges and disrupts continuity, and the potential for gaps in detected edges due to weak gradients or thresholding, leading to incomplete line segments.

Classical Methods

Hough Transform

The Hough transform, originally introduced by Paul V. C. Hough in 1962 as a method for detecting particle tracks in bubble chamber images, maps edge points from an input image to curves in a parameter space, where collinear points intersect at peaks corresponding to lines. This approach was patented in the same year and laid the foundation for line detection in computer vision. The standard form of the Hough transform, as generalized by Richard O. Duda and Peter E. Hart in 1972, parameterizes lines in polar coordinates to avoid issues with vertical lines in slope-intercept form. A line is represented by the equation \rho = x \cos \theta + y \sin \theta, where \rho is the perpendicular distance from the origin to the line, \theta is the angle of the perpendicular with respect to the x-axis, and (x, y) are coordinates of points on the line. For each edge point (x_i, y_i) in a binary edge map, the transform generates a sinusoid in the (\rho, \theta) parameter space by varying \theta and computing the corresponding \rho. An accumulator array, a 2D grid discretized over \rho and \theta, records votes by incrementing cells along each sinusoid. After processing all edge points, peaks in the accumulator—identified via local maxima detection—indicate the parameters of dominant lines, with peak height reflecting the number of supporting points. Parameter quantization is essential for implementing the accumulator, typically with \theta resolved in 1° increments over 0° to 180° and \rho in 1-pixel units scaled to the image dimensions. Finer resolutions improve accuracy but increase and , while coarser ones may miss subtle lines. To mitigate the high computational cost of the standard , which requires for every possible \theta per edge point, the probabilistic Hough transform (PHT) was developed in the early 1990s. Introduced by Nahum Kiryati, Yonina C. Eldar, and Alfred M. Bruckstein in 1991, the PHT employs random sampling by selecting pairs of edge points to define lines, computing the unique line \rho, \theta that define the line passing through the two points for each pair, and only for that parameter pair in the accumulator. This subsampling reduces the number of votes needed, often using a small fraction of edge points while maintaining detection reliability through probabilistic analysis of sampling sufficiency. The offers robustness to noise and partial occlusions, as the voting mechanism accumulates evidence from multiple collinear points, allowing lines to emerge even with gaps or outliers. However, it demands substantial memory for the accumulator array in large images, with scaling as O(\Delta \rho \times \Delta \theta), where \Delta \rho and \Delta \theta are the quantization ranges. It is typically applied to edge maps produced by prior detection steps, such as gradient-based operators.

Convolution-Based Techniques

Convolution-based techniques for line detection involve applying linear filters, or kernels, to an or its through convolution operations. These filters are specifically designed to produce high responses at pixels where lines of predetermined orientations and widths are present, effectively highlighting linear structures while suppressing noise and non-linear features. The process typically begins with preprocessing, such as using operators like Sobel or Canny, to enhance potential line locations before . This approach leverages the locality of convolution, making it computationally efficient for detecting straight lines in raster . Kernel design is central to these methods, where each is tailored to a specific line orientation θ and width. For instance, a horizontal line kernel might consist of a that emphasizes along the line direction while averaging perpendicularly, such as: \begin{bmatrix} -1 & -1 & -1 \\ 2 & 2 & 2 \\ -1 & -1 & -1 \end{bmatrix} This kernel, when convolved with the , yields positive values for bright lines on dark backgrounds; for the reverse , the signs are inverted. Kernels for other orientations, like 45° or 90°, are obtained by rotating this base mask, often using for non-axis-aligned angles. Early designs drew from domain analysis to ensure isotropic responses and optimal frequency selectivity. To handle varying line thicknesses, a multi-scale strategy is employed by using kernels of different sizes—e.g., 3×3 for thin lines and 5×5 or larger for thicker ones—or by the input prior to . Common orientations include 0°, 45°, 90°, and 135°, covering principal directions in many natural and man-made s. Response aggregation then combines outputs from multiple kernels: at each , the maximum absolute response across orientations indicates the strongest line presence, which is thresholded to produce a line map. This aggregation helps in creating a comprehensive line strength without requiring . These techniques emerged in the early , initially applied in texture analysis to segment regions based on linear primitives using quadrature filter banks. Knutsson and Granlund's work on Fourier-based laid foundational principles for rotationally invariant line detection. Advantages include high efficiency, especially on parallel hardware like GPUs, due to the separable nature of s, and simplicity in implementation. However, they are sensitive to interruptions or gaps in lines and variations in thickness, often requiring post-processing like morphological operations for robustness. The outputs can be linked to line models, such as (ρ, θ) in polar form, for further parameter estimation.

Modern Methods

Grouping and Perceptual Organization

Grouping and perceptual organization in line detection involve connecting edge pixels into coherent line segments by leveraging principles from , particularly —where aligned elements are perceived as a single unit—and proximity, where nearby elements are grouped together. These principles guide algorithms to form finite line segments from fragmented edge maps, mimicking human to identify meaningful structures in images without relying on global voting mechanisms. A prominent example is the , introduced in 2010, which performs region growing on pixels ordered by to build line segments. Starting from outputs, LSD selects seeds as pixels with high and level-line angle alignment. It then grows regions by including neighboring pixels that satisfy and proximity criteria, constrained by direction and thresholds derived from rather than fixed parameters. Once a region is grown, an elliptical kernel is fitted to validate the segment's straightness, ensuring it aligns with perceptual continuity. fitting approximates the segment's boundaries, followed by merging with adjacent segments if they share similar orientations and endpoints within a small . Segment validation in LSD employs the Number of False Alarms (NFA) statistic, an a contrario measure rooted in Gestalt-inspired , to control false positives. The NFA estimates the expected number of random alignments under a of noise; a segment is accepted if \log(\text{NFA}) < \log(\alpha), where \alpha is a small level (typically around 0.0001), indicating the detection is unlikely due to chance. This parameter-free approach avoids manual tuning, making LSD robust across diverse images. Another method is the Fast Line Detector (FLD), proposed in 2014, which uses a top-down smaller on regions of support derived from pixels in the map to extract and validate line segments. FLD identifies potential line supports by analyzing the smaller eigenvalues of local matrices, enabling accurate detection with controlled false alarms and no parameter tuning, particularly efficient for urban scenes. Both and FLD offer advantages in being largely parameter-free and capable of real-time performance, with achieving subpixel accuracy at linear O(1) per . Adaptations like Mobile (M-LSD) in the 2020s extend these to resource-constrained devices, enabling applications in mobile for on-device scene understanding.

Learning-Based Approaches

Learning-based approaches to line detection have marked a significant shift from classical methods by leveraging convolutional neural networks (CNNs) and subsequent architectures to learn hierarchical features directly from data, enabling end-to-end prediction of lines or segments without explicit or parameter space . A seminal work in this domain is the Wireframe Parsing model introduced in 2019, which builds on holistically-nested (HED) to first predict junction heatmaps and then infer line segments connecting them, achieving robust detection in structured man-made scenes. This approach treats line detection as a wireframe parsing task, outputting vectorized representations that capture global structural cues, and demonstrates superior performance over traditional baselines like the on cluttered images. Subsequent advancements incorporated learned voting mechanisms inspired by classical techniques, as seen in the Deep Hough Transform (DHT) framework from 2020, which uses a CNN backbone to generate feature maps and a differentiable Hough module for voting in dual parameter space (position and angle), allowing end-to-end training for semantic line detection. DHT excels in handling partial occlusions and varying scales by learning adaptive voting weights, reporting up to 10% improvements in average precision on benchmarks compared to prior CNN-based methods. Building on this, transformer-based models like the Line Transformer (LETR) in 2021 introduced global attention mechanisms to model long-range dependencies among potential line segments, eliminating the need for explicit edge detection and directly regressing segment endpoints from image features, which enhances accuracy in complex scenes with intersecting lines. These methods are typically trained and evaluated on specialized datasets such as the Wireframe dataset, comprising 5,000 images of man-made environments with pixel-precise annotations for junctions and lines, paired with 462 test images. is assessed using metrics like structural average precision (), which evaluates both localization accuracy and structural integrity by matching predicted segments to via alignment, emphasizing recall at high precision thresholds (e.g., [email protected]:0.95). Learning-based approaches offer key advantages, including superior handling of occlusions, varying lighting, and complex geometries through data-driven , often outperforming classical methods by 15-20% in sAP on real-world benchmarks. However, they require large labeled datasets for —such as the Wireframe set—and are computationally intensive, with times ranging from 20-50 per image on GPUs, limiting deployment on resource-constrained devices. As of 2025, emerging trends integrate learning-based line detection with multimodal for line lifting, fusing image-based wireframes with depth or multi-view via neural fields to reconstruct volumetric structures, as exemplified by the NEAT framework that distills wireframes from predictions using attraction fields and bipartite matching for junction alignment. Recent advancements include deformable transformer models like DT-LSD (2025), which support cross-scale interactions for faster and more accurate detection in diverse scenes. This enables applications in scene understanding and , where detections are lifted to manifolds.

Implementation Aspects

Algorithmic Steps and Optimization

The implementation of line detectors follows a generic pipeline that ensures robustness across diverse images. Preprocessing begins with converting the color image to grayscale to focus on intensity variations, followed by noise reduction using Gaussian blur to smooth artifacts while preserving edges. This step mitigates sensitivity to minor fluctuations, as demonstrated in standard Hough-based workflows. Subsequent edge detection identifies candidate pixels for line formation, commonly employing the Canny operator, which applies Gaussian smoothing, gradient computation via Sobel filters, non-maximum suppression, and hysteresis thresholding to yield thin, connected edges. Line fitting then aggregates these edges into parametric representations, such as using the Hough transform for infinite lines or region-growing in LSD for finite segments. Post-processing refines outputs through non-maximum suppression to eliminate overlapping detections, often based on line segment overlap thresholds to retain the strongest candidate per group. Optimization techniques address computational bottlenecks inherent in these steps. For Hough-based methods, GPU acceleration parallelizes accumulator updates, achieving significant speedups on commodity by distributing voting across thousands of threads, particularly effective for large parameter spaces. In the Probabilistic Hough Transform (PHT), hierarchical sampling reduces complexity from O(N²) in naive implementations—where N is the number of edge points and accumulator bins—to near-linear by progressively points and low-probability accumulators early. The Line Segment Detector (LSD), in contrast, achieves overall complexity through efficient region growing and validation without exhaustive voting, processing images in linear time relative to pixel count. Real-world challenges, such as varying line thicknesses and resolutions, are handled via multi-scale , where detection operates on pyramids or adaptive Gaussian scales to capture both fine and coarse structures without parameter retuning. Performance is evaluated using precision-recall curves for line-level accuracy, measuring the fraction of true positives among retrieved lines at varying thresholds, and Intersection over Union () for segment-level overlap, where thresholds like 0.5 determine matches between predicted and ground-truth segments. These metrics are benchmarked on datasets such as York Urban, comprising 102 calibrated urban with hand-labeled line segments, enabling standardized assessment of detection quality in indoor and outdoor scenes.

Practical Examples and Code Snippets

Line detection algorithms find widespread application in tasks such as autonomous driving, where they identify lane markings on roads to assist vehicle . For instance, the is commonly used to detect straight lines in images of road scenes, enabling processing at rates exceeding 30 frames per second on standard hardware. In document analysis, line detection helps in correction for scanned pages by identifying text baselines, improving accuracy. Another practical example is in , where line detection segments edges in camera feeds to overlay virtual objects aligned with real-world structures, as demonstrated in mobile AR frameworks achieving sub-pixel precision. For industrial inspection, convolution-based methods like the followed by line fitting detect defects in manufactured parts, such as cracks in metal sheets, with high detection rates in controlled lighting conditions.

Example in

A basic implementation of the probabilistic in using processes an input image to detect line segments. The algorithm first applies with Canny, then accumulates votes in parameter space to identify prominent lines, thresholding to filter weak candidates. This approach is efficient for sparse line scenes, reducing compared to the standard , which is O(n^2), through efficient sampling.
python
import cv2
import [numpy](/page/NumPy) as np

# Load image and convert to grayscale
image = cv2.imread('input_image.jpg')
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

# Apply Canny edge detection
edges = cv2.Canny(gray, 50, 150, apertureSize=3)

# Probabilistic Hough Transform
lines = cv2.HoughLinesP(edges, 1, np.pi/180, threshold=100, minLineLength=100, maxLineGap=10)

# Draw detected lines on original image
for line in lines:
    x1, y1, x2, y2 = line[0]
    cv2.line(image, (x1, y1), (x2, y2), (0, 255, 0), 2)

cv2.imshow('Detected Lines', image)
cv2.waitKey(0)
cv2.destroyAllWindows()
This code snippet, adapted from documentation, outputs an image with green lines overlaying detected segments, suitable for applications like detection.

Line Segment Detector () Example

The algorithm, a gradient-based method, detects straight without predefined thresholds, making it robust to noise and suitable for natural images. It performs region growing on edge maps and validates segments via Helmholtz principle, showing good repeatability in benchmarks compared to Hough variants. In , can be implemented via OpenCV's built-in function, which processes the in linear time relative to count. Here's a concise example for detecting lines in a cluttered scene:
python
import cv2
import [numpy](/page/NumPy) as np

# Load [image](/page/Image)
image = cv2.imread('input_image.jpg')
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

# Apply [LSD](/page/LSD)
lsd = cv2.createLineSegmentDetector(0)
lines = lsd.detect(gray)[0]  # Returns (N,1,4) array: [x1,y1,x2,y2]

# Draw lines
drawn_image = image.copy()
if lines is not None:
    for line in lines:
        x1, y1, x2, y2 = line[0]
        cv2.line(drawn_image, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 1)

    # Refine and visualize
    drawn_lines = lsd.drawSegments(drawn_image, lines)

cv2.imshow('LSD Lines', drawn_image)
cv2.waitKey(0)
cv2.destroyAllWindows()
This implementation highlights 's ability to extract perceptually meaningful segments, as validated in benchmarks showing fewer false positives than Sobel-based methods.

Learning-Based Line Detection with

Modern approaches employ convolutional neural networks (CNNs) for end-to-end line detection, such as Wireframe Parsing Networks, which predict junction-line maps from RGB images. Trained on synthetic and real datasets, these models achieve state-of-the-art performance on datasets like YorkUrbanDB, outperforming classical methods in urban scenes. A PyTorch-based snippet for inference using a pre-trained model like L-CNN illustrates heatmap for line proposals, followed by non-maximum suppression. This is particularly useful in for obstacle boundary detection. (Note: Load the model from the official repository at https://github.com/zhou13/lcnn following its instructions.)
python
import torch
import cv2
import [numpy](/page/NumPy) as np

# Load pre-trained model from official L-CNN repository (https://github.com/zhou13/lcnn)
# Example: model = load_model('path_to_pretrained.pth')  # Follow repo instructions
# model.eval()

# Load and preprocess image
image = cv2.imread('input_image.jpg')
input_tensor = [torch](/page/Torch).from_numpy(image.transpose(2,0,1)).float().unsqueeze(0) / 255.0

# [Inference](/page/Inference) (simplified)
# with [torch](/page/Torch).no_grad():
#     outputs = model(input_tensor)
#     lines = post_process(outputs)  # To get line endpoints

# Draw lines (simplified)
result = image.copy()
# for line in lines:
#     x1, y1, x2, y2 = line
#     cv2.line(result, (int(x1), int(y1)), (int(x2), int(y2)), (255, 0, 0), 2)

# cv2.imshow('Deep Lines', result)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
This example leverages for performance, with the model processing 640x480 images at around 15 on a GPU, as reported in the original implementation.

References

  1. [1]
  2. [2]
    US3069654A - Method and means for recognizing complex patterns
    Application of the hough transform to modeling the horizontal component of road geometry and computing heading and curvature. US20040086082A1 * 2002-11-05 ...
  3. [3]
    LSD: A Fast Line Segment Detector with a False Detection Control
    Dec 31, 2008 · We propose a linear-time line segment detector that gives accurate results, a controlled number of false detections, and requires no parameter tuning.
  4. [4]
    Line Segment Detection and Refinement with Deep Image Gradients
    Dec 15, 2022 · Our new line segment detector, DeepLSD, processes images with a deep network to generate a line attraction field, before converting it to a surrogate image ...
  5. [5]
    A Comprehensive Review of Image Line Segment Detection and ...
    Apr 29, 2023 · This study fills the gap by comprehensively reviewing related studies on detecting and describing two-dimensional image line segments.
  6. [6]
    [PDF] Line Segment Detection Using Transformers Without Edges
    Typically, line (segment) detection performs edge detection [3, 23, 7, 8, 32], followed by a perceptual grouping [13, 27, 10] process.
  7. [7]
    Line Segment Detection Using Transformers without Edges
    ### Summary of https://arxiv.org/abs/2101.01909
  8. [8]
  9. [9]
    A statistical method for line segment detection - ScienceDirect.com
    Line segment detection is a fundamental procedure in computer vision, pattern recognition, or image analysis applications. This paper proposes a statistical ...
  10. [10]
    Progress in Picture Processing: 1969--71 | ACM Computing Surveys
    GRIFFITH, A K "Mathematical models for automatic line detection " In {33~, pp 17-26 ... D "Edge detection in pictures by computer using planning " In {34} ...
  11. [11]
    Mathematical Models for Automatic Line Detection
    A particular decision-theoretic approach to the problem of detecting straight edges and lines in pictures is discussed. A model is proposed of the ...
  12. [12]
    Self-supervised Occlusion-aware Line Description and Detection
    Apr 7, 2021 · SOLD2 is a self-supervised method for joint detection and description of line segments, using a deep network and robust to occlusions.
  13. [13]
    A comprehensive study on lane detecting autonomous car using ...
    Dec 15, 2023 · Lane Detection is another prominent method deployed in self-driving cars, allowing it to follow the particular lane and detect turns and curves ...
  14. [14]
    Lane-GAN: A Robust Lane Detection Network for Driver Assistance ...
    Lane line detection algorithms based on deep learning can automatically extract the lane line features without making assumptions about the road structure and ...
  15. [15]
    Skew detection and correction in document images based on ...
    In this paper, a novel skew detection method based on straight-line fitting is proposed.
  16. [16]
    A Novel Adaptive Deskewing Algorithm for Document Images - PMC
    Oct 18, 2022 · The line detection-based method and the projection profile-based method have high accuracy for estimating skew angle on text documents. The ...
  17. [17]
    [PDF] Line Drawing Interpretation in a Multi-View Context - Inria
    Apr 9, 2015 · Our work combines the strength of these two domains to interpret line drawings of imaginary objects drawn over photographs of an existing scene.
  18. [18]
    [PDF] Semantic Interpretation of Architectural Drawings - CumInCAD
    Abstract. The paper reviews the needs and issues of automatically interpreting architec- tural drawings into building model representations.
  19. [19]
    PCA‐Based Line Detection from Range Data for Mapping and ...
    Apr 4, 2017 · This paper presents an original technique for robust detection of line features from range data, which is also the core element of an ...<|separator|>
  20. [20]
    Multi-Robot Collaborative Mapping with Integrated Point-Line ... - NIH
    Sep 4, 2024 · This paper proposes a multi-robot collaborative mapping method based on point-line fusion to address this issue.
  21. [21]
    Block-based Vanishing Line and Vanishing Point Detection for 3D ...
    This paper presents a robust and reliable block-based method for vanishing line and vanishing point detection. The proposed method provides assistance in ...
  22. [22]
    Horizon Line Detection in Historical Terrestrial Images in ... - MDPI
    A robust and accurate approach for horizon line detection in historical monochrome images in mountainous terrain was developed.
  23. [23]
    History Of Computer Vision - Let's Data Science
    1980 : NeoCognitron ... The Neocognitron is a neural network model for pattern recognition and image processing, developed by Kunihiko Fukushima in the 1980s.
  24. [24]
    Lane Line Map Estimation for Visual Alignment - IEEE Xplore
    This paper presents a novel lane line map estimation method from single images, which is applicable for visualization tasks such as augmented reality (AR) ...
  25. [25]
    Improving the Classification Accuracy of Accurate Traffic Sign ...
    Aug 7, 2025 · Results showed that the combined features of HOG and LBP offers the best classification performance, and the recognition rate is over 93.6%.
  26. [26]
    [PDF] Use of the Hough Transformation To Detect Lines and Curves in ...
    The parameter space is defined by the parametric representation used to describe lines in the picture plane. Hough chose to use the familiar slope-intercept ...
  27. [27]
    None
    Nothing is retrieved...<|separator|>
  28. [28]
    [PDF] Structure and motion from straight line segments
    Structure and motion from straight line segments. J.M.M. Montiel*, J.D. ... Computer Vision. His current research interests include sensor integration ...
  29. [29]
    [PDF] Chapter 5. Edge Detection
    Edge detection identifies significant local changes in image intensity, often at boundaries between regions, and is the first step in recovering image ...
  30. [30]
    [PDF] A Review of Classic Edge Detectors - IPOL Journal
    Jun 16, 2015 · In this paper some of the classic alternatives for edge detection in digital images are studied. The main idea of edge detection algorithms is ...
  31. [31]
    A Computational Approach to Edge Detection - IEEE Xplore
    This paper describes a computational approach to edge detection. The success of the approach depends on the definition of a comprehensive set of goals.
  32. [32]
    Use of the Hough transformation to detect lines and curves in pictures
    Jan 1, 1972 · Hough, P.V.C. Method and means for recognizing complex patterns. U.S. Patent 3,069,654, Dec. 18, 1962. Google Scholar. [3]. Griffith, A.K. ...
  33. [33]
    [PDF] A Survey on Hough Transform, Theory, Techniques and Applications
    Speedup, Memory Saving. 1. Introduction. In 1962 Paul Hough introduced an efficient method for detecting lines in binary images [1] ...
  34. [34]
    A probabilistic Hough transform - ScienceDirect.com
    The performance of the resulting “Probabilistic Hough Transform” is analysed. The analysis is supported by experimental evidence. Previous article ...
  35. [35]
    Line Detection
    The line detection operator consists of a convolution kernel tuned to detect the presence of lines of a particular width n, at a particular orientation.
  36. [36]
  37. [37]
    (PDF) Gestalt Theory and Computer Vision - ResearchGate
    Gestalt perception contains a series of laws to explain human perception, such as laws of similarity, proximity, closeness, smoothness, symmetry and so on.Missing: collinearity | Show results with:collinearity
  38. [38]
  39. [39]
    [PDF] a Line Segment Detector - LSD - IPOL Journal
    LSD is a linear-time Line Segment Detector giving subpixel accurate results. It is designed to work on any digital image without parameter tuning.
  40. [40]
    [1905.03246] End-to-End Wireframe Parsing - arXiv
    May 8, 2019 · We present a conceptually simple yet effective algorithm to detect wireframes in a given image. Compared to the previous methods which first predict an ...
  41. [41]
    Deep Hough Transform for Semantic Line Detection - arXiv
    Mar 10, 2020 · In this paper, we incorporate the classical Hough transform technique into deeply learned representations and propose a one-shot end-to-end learning framework ...
  42. [42]
    A line-segment-based non-maximum suppression method for ...
    Sep 5, 2022 · This paper proposes a novel and simple method by utilizing the distribution of line segments to facilitate the Non-Maximum Suppression (NMS) for the object ...
  43. [43]
    Fast Hough Transform on GPUs: Exploration of Algorithm Trade-Offs
    This paper introduces two new implementations of the Hough transform for lines on a GPU. One focuses on minimizing processing time, while the other has an input ...
  44. [44]
  45. [45]
    [PDF] Multiscale Detection of Curvilinear Structures in 2-D and 3-D Image ...
    This paper presents a novel, parameter-free tech- nique for the segmentation and local description of line structures on multiple scales, both in 2-D and 3-D.
  46. [46]
    York Urban Line Segment Database Information - Elder Laboratory
    The York Urban Line Segment Database is a compilation of 102 images (45 indoor, 57 outdoor) of urban environments consisting mostly of scenes from the campus ...