A Review of Image Segmentation Evaluation in the 21st Century

A Review of Image Segmentation Evaluation in the 21st Century

Yu-Jin Zhang (Department of Electronic Engineering, Tsinghua University, China)
Copyright: © 2015 |Pages: 11
DOI: 10.4018/978-1-4666-5888-2.ch579
OnDemand PDF Download:
List Price: $37.50

Chapter Preview



As mentioned above, most empirical evaluation methods can be classified into goodness method group and discrepancy method group (Zhang 1996). The goodness method can perform the evaluation without the help of reference images while the discrepancy method needs some reference images to arbitrate the quality of segmentation. More importantly, they use different empirical criteria for judging the performance of segmentation algorithms. These criteria play some critical roles in determining the generality, usability, sensitivity, effectiveness and efficiency of these evaluation procedures.

In Table 1, the already reviewed and compared empirical criteria for image segmentation evaluation are summarized (Zhang, 1996; Zhang, 2001). In Table 1, three groups of criteria can be distinguished: G for goodness criteria, D for discrepancy criteria and S for specialized criteria. The criteria for the last group have some particularity so as to be different from either goodness criteria or discrepancy criteria, but the methods using these criteria can still be classified as goodness like or discrepancy like ones.

Table 1.
A list of empirical criteria for evaluation and their method groups
Criterion GroupNo.Criterion NameMethod Class
G-1Intra-region uniformityGoodness
G-2Inter-region contrastGoodness
G-3Region shapeGoodness
G-4Moderate number of regionsGoodness
D-1Number of mis-segmented pixelsDiscrepancy
D-2Position of mis-segmented pixelsDiscrepancy
D-3Number of objects in the imageDiscrepancy
D-4Feature values of segmented objectsDiscrepancy
D-5Miscellaneous object quantitiesDiscrepancy
D-6Region consistencyDiscrepancy
D-7Grey level differenceDiscrepancy
D-8Symmetric divergence (cross-entropy)Discrepancy
S-1Amount of editing operationsDiscrepancy like
S-2Visual inspectionDiscrepancy like
S-3Correlation between original image and bi-level imageGoodness like

Key Terms in this Chapter

Image Segmentation: A process consists of subdividing an image into its constituent parts and extracting these parts of interest (objects) from the image.

Edge Detection: Edge detection is the most common approach for detecting discontinuities in images, and is the fundamental step in edge-based parallel process for segmentation. An edge is a local concept. To form a complete boundary of an object, edge detection should be followed by edge linking or connection.

Watersheds: Watershed technique is inspired from the topographic interpretation of image segmentation by watersheds embodies many concepts of edge detection, thresholding and region processing techniques, and often produces stable and continuous results.

Graph Search: Graph search is a particular type of segmentation technique which combines edge detection and linking together. It represents edge segments in the form of a graph and searches the graph for low-cost paths that correspond to significant edges or boundaries of objects.

Active Contour Model: Active contour model is a sequential technique for image segmentation. Given an approximation of the boundary of an object in an image, an active contour model can be used to find the actual boundary by deforming the initial boundary to lock onto features of interest within this image.

Image Engineering: Image engineering is an integrated discipline/subject comprising the study of all the different branches of image and video techniques. It mainly consists of three levels: image processing, image analysis, and image understanding.

Gradient Operator: Gradient operator is the first type of operator used for edge detection. The gradient of an image is a vector consisting of the first order derivatives (including the magnitude and direction) of an image.

Clustering: Clustering is also called unsupervised learning and is a powerful technique for pattern classification. It is a process to group, based on some defined criteria, two or more terms together to form a large collection In the context of image segmentation, it is often considered as the multi-dimensional extension of the thresholding technique.

Thresholding: Thresholding techniques are the most popularly used segmentation techniques. A set of suitable thresholds need to be first determined, and then the image can be segmented by comparing the pixel properties with these thresholds.

Region Growing: Region growing is a region-based sequential technique for image segmentation by assembling pixels into larger regions based on predefined seed pixels, growing criteria, and stop conditions.

Complete Chapter List

Search this Book: