Knowledge modeling for the image understanding task as a design task

https://doi.org/10.1016/j.eswa.2005.01.020Get rights and content

Abstract

This paper formally analyzes the image understanding (IU) task at the knowledge level and in the observer domain. The analysis is done at three levels: task, method and domain knowledge, distinguishing the generic components in most of the IU tasks, thereby enabling the components to be reused. We model the IU problem as a design task and define a generic problem solving method (PSM) that allows us to tackle the task in a hierarchical and recursive way. The main advantage of this generic PSM is the possibility of instantiating specific PSMs through parameter space configuration, which enables the structure to be reused both across the task decomposition at different hierarchical levels and across different application domains. This work has been done following the well-established foundations of knowledge engineering that prescribe the maintenance of the conceptual structure from the modeling stage at the knowledge level down to the implementation. Finally, we apply the proposed framework to the problem of identifying the papilla in eye fundus images in order to exemplify the successive stages in the modeling process and system design, and accordingly justify the framework's validity.

Introduction

It is widely accepted that the IU task is a complex task requiring knowledge of the application domain for its solution (Crevier and LePage, 1997, Rao and Jain, 1988, Tsotsos, 1987). In fact, many IU systems use domain knowledge explicitly. These systems can be classified into two large groups: (1) Generic systems, with sufficient flexibility and capacity for representation so as to be able to integrate any type of knowledge within the system (Draper et al., 1989, Matsuyama and Hwang, 1990); and (2) more specific systems, based on assumptions that simplify the problem, such as starting from a prior segmentation focusing on the objective (Brooks, 1983, Grimnes and Aamodt, 1996, Kummert et al., 1993) or on the assumption that there are some domain structure properties that make it possible to define a method capable of navigating in this structure (Chan et al., 1996, Gong and Kulikowski, 1995, Kummert et al., 1993, Strat and Fischler, 1991). This simplification can be taken to the limit and generate a specific application for a particular domain (Cabello et al., 1999, McKeown et al., 1985).

The main criticism of generic systems is that, as regards flexibility, a large part of the problem is transferred to the system designer, which must program both the domain knowledge and the method used. These architectures do not help at the system definition stage, since they only make computable that knowledge to be explicitly used in accordance with the inferential strategy. Moreover, the systems in the second group are capable of proposing specific solutions, but because they do not differentiate the distinct types of knowledge used, these solutions cannot be extrapolated to other domains with slightly different initial assumptions. Finally, the main problem detected in both cases is that the analysis of the knowledge has focused directly on the representation at the symbolic level instead of modeling first at the knowledge level.

Additionally, in the last two decades the usefulness of reusing modeling components at the knowledge level has been shown (Chandrasekaran, 1986, Chandrasekaran, 1990, Clancey, 1985, Schreiber et al., 1999, Schreiber et al., 1994). The main advantage is that it has been possible to understand the problem of knowledge engineering as a modeling activity, where there is an initial structure in the method that solves the task, which has to be linked to the knowledge in the application domain by assigning this knowledge to method roles. It thus becomes evident what the purpose of the domain knowledge is for resolving a task: it guides the dialogue with the domain expert for the acquisition of his knowledge at the design stage and later facilitates the system maintenance and documentation and component reuse. This work tries to resolve some of the difficulties encountered in the IU task, modeling at knowledge level the use of domain knowledge. Thus a modeling framework is defined that facilitates the development of IU systems for any domain.

At the knowledge level, the IU task allows two basic formulations: as analysis or synthesis tasks. It is advisable to model IU as an analysis task when it is possible to associate certain pixel configurations with objects of interest in the image, whereas the IU must be modeled as a synthesis task when the solution has to be constructed and not selected. Because of the uncertainty associated with the vision process, the most usual case and also the one followed in this work, is to consider the IU task as a synthesis task. At the knowledge level, we decompose the task following the CommonKads Expertise Model. The proposed PSM represents a common framework for the integration of different PSMs in the solution of complex problems. This is fundamental, because we start from the actual fact that a complex task, like IU, cannot usually be resolved with just one PSM, since each PSM introduces a bias at the knowledge acquisition stage restricting the type of knowledge that can be used in inference. Accordingly, it is necessary to leave open the possibility of using different PSMs depending on the type of knowledge involved. Thus it is the domain knowledge that makes it possible to select the right PSM at every moment, without imposing a prior structure for the domain knowledge. This idea is in the line of the competence theory proposed by Wielinga (Wielinga, Akkermans, & Schreiber, 1995), which is different from that developed by Löckenhoff and Messer in (Breuker, 1994), where in the analysis of the problem there are different approaches according to the domain knowledge application type: case-based methods, resource-based methods and structure-based methods. For real problems it is generally not possible to have such a clear division of the knowledge available and, therefore, the solution method is usually a hybrid method.

To illustrate the different modeling stages, we are going to show the results obtained in the medical field, where the form of data acquisition, deformable nature of the objects, statistical variation of normal and abnormal information are characteristics (Duncan & Adache, 2000) that make it especially suitable for this methodology. In particular, we focus on the application for identifying the optic nerve head (ONH) from eye fundus color images.

In Section 2 of this article the IU task at the knowledge level is analyzed and the task ontology for IU as a routine design task is described. In Section 3 the design task solution is characterized as a search process in the design space and the structure of a generic PSM is described. Finally, to exemplify and validate the proposed framework, Section 4 describes its application to identifying the optic nerve head.

Section snippets

The image understanding task

We analyze the IU task as a routine design task, the aim of which is to construct a system or design from a predefined set of basic components, constraints, requirements and preferences. Thus we construct the solution during the process due to the uncertainty associated with the knowledge available in each task and we assume that all the elements used to construct the solution are known in advance. Then we establish the following correspondence between the design task entities and the IU task

A generic method for routine design problem solving

Routine design problem solving has been characterized as a search process in a design space (Chandrasekaran, 1990, Marcus, 1988, Motta, 1999, Stefik, 1995, Wielinga et al., 1995, Wielinga and Schreiber, 1997). This work has generalized a proposal previously presented by Motta (1999) for parametric design. This proposal is based on defining a generic PSM for the task and independent of the specific method so that it acts as a reference for describing the different types of knowledge used for

Application to the location of the optic nerve head

In this section we are going to illustrate the design process stages. The ultimate aim of the task is to identify the optic nerve head in eye fundus color images for its use in diagnosis tasks. However, to illustrate the method, this article concentrates solely on locating the Papilla.

Conclusions

This work has developed a knowledge model for the IU task, distinguishing: task, method and application domain knowledge. To resolve the task a generic PSM was chosen because the IU task is a complex task that is not resolved with just one PSM, since each PSM introduces a bias at the knowledge acquisition stage restricting the type of knowledge that can be used in inference. At the knowledge level, this PSM represents a common framework for integrating different PSMs in the solution of complex

Acknowledgements

We would like to thank to the Opthtalmology Service of Miguel Servet Hospital and to the Oftalmology Department of the San Carlos Clinic Hospital for letting us using their images in this study. We would like to thank the CICYT for its support, through the DIAGEN project (TIC-97-0604), in the context of which this case study has been carried out.

References (28)

  • W. Clancey

    Heuristic classification

    Artificial Intelligence

    (1985)
  • D. Crevier et al.

    Knowledge-based image understanding systems: A survey

    Computer Vision and Image Understanding

    (1997)
  • T.R. Gruber et al.

    The configuration design ontologies and the VT elevator domain theory

    International Journal of Human-Computer Studies

    (1996)
  • F. Kummert et al.

    Control and explanation in a signal understanding environment

    Signal Processing

    (1993)
  • J. Breuker

    CommonKads lybrary for expertise modelling

    (1994)
  • R. Brooks

    Model-based three-dimensional interpretations of two-dimensional images

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (1983)
  • D. Cabello et al.

    Automatic segmentation of lung fields on chest radiographic images

    Computeres and Biomedical Research

    (1999)
  • S. Chan et al.

    Object-oriented knowledge-based system for image diagnosis

    Applied AI

    (1996)
  • B. Chandrasekaran

    Generic tasks in knowledge based reasoning: High level building blocks for expert system design

    IEEE Expert

    (1986)
  • B. Chandrasekaran

    Design problem solving: A task analysis

    AI Magazine

    (1990)
  • B.A. Draper et al.

    The schema system

    International Journal of Computer Vision

    (1989)
  • J.S. Duncan et al.

    Medical image analysis: Progress over two decades and the challenges ahead

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2000)
  • L. Gong et al.

    Composition of image analysis processes through object-centered hierarchical planning

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (1995)
  • M. Grimnes et al.

    A two layer case-based reasoning architecture for medical image understanding

    (1996)
  • Cited by (7)

    • Modeling spatial layout for scene image understanding via a novel multiscale sum-product network

      2016, Expert Systems with Applications
      Citation Excerpt :

      Natural scene image understanding, which aims at labeling each pixel of a scene image to a predefined object class and simultaneously performing segmentation or recognition of multiple objects that occur in the scene, has been extensively studied in the past years (Farabet, Couprie, Najman, & LeCun, 2013; Gritti, Damkat, & Monaci, 2013; Liu, Xu, & Feng, 2011; Rincón, Bachiller, & Mira, 2005; Shotton, Winn, Rother, & Criminisi, 2009; Tighe & Lazebnik, 2013; Tu, Chen, Yuille, & Zhu, 2005; Yin, Jiao, Chai, & Fang, 2015).

    • Segmentation of Glaucoma Disease based on Modified Kernel Fuzzy C-Means Algorithm

      2021, Proceedings of the 2nd International Conference on Electronics and Sustainable Communication Systems, ICESC 2021
    • The use of Knowledge in Biomedical Image Analysis

      2015, Biomedical Image Understanding: Methods and Applications
    • ARDIS: Knowledge-based dynamic architecture for real-time surface visual inspection

      2009, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    View all citing articles on Scopus
    View full text