Source-linked AI summary
Image Segmentation Algorithms Overview
Song Yuheng, Yan Hao
TL;DR
Image segmentation is widely used, but its methods are diverse and lack a recognized unified standard. This paper analyzes and compares threshold, regional growth, edge detection, clustering, and weakly supervised CNN approaches. It concludes that combining multiple segmentation methods can improve segmentation effects, while no single method adapts to all cases.
Problem
Image segmentation methods are diverse and complex, with no recognized unified standard, motivating comparison of their advantages and disadvantages.
Method
The paper analyzes and compares multiple image segmentation approaches, including region-based, edge detection, clustering, and weakly supervised CNN methods.
Results
Combining multiple segmentation methods can improve the segmentation effect.
Takeaways & Limitations
Future image segmentation development is forecast to involve combinations of multiple segmentation methods.
Takeaways & Limitations
No single segmentation method can adapt to all cases.
Abstract
from arXiv · showhide
The technology of image segmentation is widely used in medical image processing, face recognition pedestrian detection, etc. The current image segmentation techniques include region-based segmentation, edge detection segmentation, segmentation based on clustering, segmentation based on weakly-supervised learning in CNN, etc. This paper analyzes and summarizes these algorithms of image segmentation, and compares the advantages and disadvantages of different algorithms. Finally, we make a prediction of the development trend of image segmentation with the combination of these algorithms.
1 Introduction
Image segmentation divides an image into meaningful regions and supports image analysis, feature extraction, and recognition. This paper reviews several segmentation approaches, compares their advantages and disadvantages, and considers combining methods for future solutions.
- Image segmentation divides an input image into regions with the same category or characteristics to extract areas of interest.
- The paper analyzes threshold, regional growth, edge detection, clustering, and weakly supervised CNN segmentation methods.
- Threshold segmentation: Threshold segmentation selects an optimal threshold according to a criterion and groups pixels by gray level.
- Regional growth segmentation: Regional growth begins with seed pixels and merges neighboring pixels with similar properties into regions.
- Edge detection and clustering: Edge detection uses gray-level or color discontinuities to identify boundaries, while clustering divides samples by similarity.
- The paper notes that segmentation procedures are diverse and complex, lack a recognized unified standard, and require comparison of methods and future solution analysis.
2.1 Region-based Segmentation
Region-based segmentation includes thresholding and regional growth. Thresholding is simple and fast but sensitive to weak contrast, noise, and uneven grayscale, while regional growth can provide connected regions and boundary information but depends on seeds and growth criteria.
- Threshold segmentation: Threshold segmentation divides images into target and background regions using gray-level thresholds, either globally or locally.
- Threshold segmentation: Otsu thresholding selects a globally optimal threshold by maximizing the variance between classes.
- Threshold segmentation: Threshold methods are simple and fast, and can segment targets effectively when target-background contrast is high.
- Threshold segmentation: Threshold segmentation is less accurate when grayscale differences are weak or gray values overlap, and it is sensitive to noise and unevenness because spatial information is omitted.
- Regional growth segmentation: Regional growth selects seed pixels and merges neighboring pixels when their gray-value difference from the seed is below threshold T.
- Regional growth segmentation: At T = 3 the example is divided into two regions, whereas at T = 6 the whole plot becomes one area, showing the importance of threshold choice.
- Regional growth segmentation: Regional growth can separate connected regions with similar characteristics and provide boundary information using few seed points and flexible criteria.
- Regional growth segmentation: Regional growth has high computational cost and can produce voids or over-division under noise and grayscale unevenness, with shadows also affecting images.
2.2 Edge Detection Segmentation
Edge-detection segmentation identifies image boundaries through local intensity discontinuities and differential operators. The section describes first- and second-order operators, emphasizing Sobel’s noise reduction and Laplacian’s sensitivity to noise.
- Image edges correspond to local discontinuities such as gray-value, color, or texture changes.Differential operations detect these discontinuities to support segmentation.
- Spatial-domain differential operators convolve templates with images, and edge detection is generally used as preprocessing.Prewitt, Roberts, and Sobel are common first-order operators; Laplacian, Kirsch, and Wallis are second-order examples.
- Sobel Operator: Sobel’s local averaging can smooth noise and reduce its influence during edge detection.The section also reports that Sobel performs better than Prewitt and Roberts because pixel positions are weighted.
- Sobel Operator: Sobel estimates horizontal and vertical gradients with two 3x3 templates and combines them to calculate gradient magnitude.Its local averaging smooths noise and gives it an advantage over Prewitt and Roberts operators.
- Laplacian Operator: The Laplacian can combine with smoothing to form a new template and is used to address image blurring associated with imaging diffusion.The section states that it is suitable for noise-free images and should be preceded by low-pass filtering when noise is present.
- Laplacian Operator: The Laplacian is an isotropic, rotationally invariant second derivative that responds more strongly to isolated pixels than to edges or lines.Because noise produces strong isolated-pixel responses, Laplacian edge detection requires low-pass filtering in noisy images.
2.3 Segmentation based on clustering
Clustering-based segmentation maps image pixels into feature space, groups similar points there, and maps the groups back to the image. K-means is presented as a fast, simple, efficient, and scalable clustering method.
- Clustering-based segmentation maps image pixels into feature space, partitions aggregated feature points, and maps the result back to image space.This approach organizes pixels according to their feature-space grouping.
- K-means: K-means groups samples into compact clusters according to distance, placing closer points into the same cluster.The method is described as one of the most commonly used clustering algorithms.
- K-means: K-means selects K initial centers, assigns samples to the nearest center, recomputes means, and repeats until convergence or an iteration limit.These four steps define the algorithm’s iterative clustering procedure.
- K-means: K-means is fast and simple, with high efficiency and scalability for large datasets.The supplied passage states these as advantages of the algorithm.
2.4 Segmentation based on weakly-supervised learning in CNN
Weakly supervised CNN segmentation uses limited annotations with DeepLab and expectation maximization to estimate unmarked pixel classes and CNN parameters. The section describes coarse prediction followed by CRF refinement and reports better results from bounding-box training than image-level labels alone.
- Weakly and semi-supervised DCNN methods apply limited annotations to semantic image segmentation.The discussion identifies image-level labels and bounding boxes as forms of markup training data.
- DeepLab Training: The described method uses DeepLab with expectation maximization to estimate unmarked pixel classes and CNN parameters.DeepLab is organized into a coarse prediction stage and a refinement stage.
- DeepLab Training: DeepLab first obtains a coarse score map with an FCN and interpolates it to the original image size.This is the first of the two described processing steps.
- DeepLab Training: The second DeepLab step uses a fully connected CRF to refine segmentation details.The CRF is introduced after coarse prediction and interpolation.
- Annotation Strategies: Bounding-box training uses CRF-generated segmentations followed by full supervision on those segmentations.This procedure is described for training images with bounding-box marks.
- Annotation Strategies: Bounding-box training produces better results than using image-level labels alone for segmentation.The section explicitly characterizes image-level-only segmentation as poor and bounding-box results as better.
3 Conclusion
Image segmentation research remains theoretically incomplete and faces practical application problems, making it difficult to find one method that adapts to all images. The paper therefore points toward combining segmentation methods and using machine learning to select parameters or define regions of interest.
- Image diversity and uncertainty make it difficult to find a segmentation method that adapts to all images.
- Image segmentation theory is not yet perfect, and applied research still contains many practical problems.
- Combining multiple segmentation methods and fusing multiple features is proposed to use different algorithms’ advantages for better segmentation.
- Machine learning can support parameter selection, including thresholds for threshold segmentation and K values for K-means.
- A CNN can frame the region of interest before applying a non-machine-learning segmentation method to improve segmentation effect.
- The paper anticipates further development and wider use of image segmentation methods.