On the use of the overlapping area matrix for image segmentation evaluation: A survey and new performance measures

University of the Balearic Islands, Department of Mathematics and Computer Science, Palma de Mallorca, Spain
Pattern Recognition Letters (Impact Factor: 1.27). 01/2006; DOI: 10.1016/j.patrec.2006.05.002
Source: DBLP

ABSTRACT The development of common and reasonable criteria for evaluating and comparing the performance of segmentation algorithms has always been a concern for researchers in the area. As it is discussed in the paper, some of the measures proposed are not adequate for general images (i.e. images of any sort of scene, without any assumption about the features of the scene objects or the illumination distribution) because they assume a certain distribution of pixel gray-level or colour values for the interior of the regions. This paper reviews performance measures not performing such an assumption and proposes a set of new performance measures in the same line, called the percentage of correctly grouped pixels (CG), the percentage of over-segmentation (OS) and the percentage of under-segmentation (US). Apart from accounting for misclassified pixels, the proposed set of new measures are intended to compute the level of fragmentation of reference regions into output regions and vice versa. A comparison involving similar measures is provided at the end of the paper.

  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: The strategic management of agricultural lands involves crop field monitoring each year. Crop discrimination via remote sensing is a complex task, especially if different crops have a similar spectral response and cropping pattern. In such cases, crop identification could be improved by combining object-based image analysis and advanced machine learning methods. In this investigation, we evaluated the C4.5 decision tree, logistic regression (LR), support vector machine (SVM) and multilayer perceptron (MLP) neural network methods, both as single classifiers and combined in a hierarchical classification, for the mapping of nine major summer crops (both woody and herbaceous) from ASTER satellite images captured in two different dates. Each method was built with different combinations of spectral and textural features obtained after the segmentation of the remote images in an object-based framework. As single classifiers, MLP and SVM obtained maximum overall accuracy of 88%, slightly higher than LR (86%) and notably higher than C4.5 (79%). The SVM+SVM classifier (best method) improved these results to 89%. In most cases, the hierarchical classifiers considerably increased the accuracy of the most poorly classified class (minimum sensitivity). The SVM+SVM method offered a significant improvement in classification accuracy for all of the studied crops compared to the conventional decision tree classifier, ranging between 4% for safflower and 29% for corn, which suggests the application of object-based image analysis and advanced machine learning methods in complex crop classification tasks.
    Remote Sensing 05/2014; 6(6):5019-5041. · 2.62 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: There are close relationships between three popular approaches to image thresholding, namely Ridler and Calvard’s iterative-selection (IS) method, Kittler and Illingworth’s minimum-error-thresholding (MET) method and Otsu’s method. The relationships can be briefly described as: the IS method is an iterative version of Otsu’s method; Otsu’s method can be regarded as a special case of the MET method. The purpose of this correspondence is to provide a comprehensive clarification, some practical implications and further discussions of these relationships.
    Pattern Recognition Letters 04/2012; 33(6):793–797. · 1.27 Impact Factor
  • Annals of Emergency Medicine - ANN EMERG MED. 01/2011; 58(4).