Defining a target distinctness measure through a single-channel computational model of vision

https://doi.org/10.1016/S0167-8655(02)00283-0Get rights and content

Abstract

The main goal of this paper is to define a computational measure for quantifying the visual distinctness of targets in complex natural scenes, using a single-channel vision model. The basic features incorporated in the defined target distinctness measure are: (i) an adaptive contrast sensitivity function; (ii) an estimation of image contrast masking and (iii) an estimation of a set of interest points in the image. Several experiments are performed to determine the relative contribution of each of these factors to the measure, using the correlation with human performance data as the evaluation function.

Introduction

Visual target acquisition is a complex process, and many factors involved are not yet fully understood. In fact, targets which are similar to their background or to many structures in other parts of the scene are harder to detect than targets which are highly dissimilar to these structures. Also, the visual distinctness of a target decreases when increasing the variability of the scene. Visual distinctness measures are used to compare and evaluate target visibility, and to quantify the complexity of a scene. Several studies show that simple quantitative image measures do not give good predictive results when are applied to the distinctness of targets in complex background scenes. For example, Saghri et al. (1989) and Daly (1993) show the severe failure of a common error metric: the mean square error (MSE). Another common quantitative metric is the root mean square error (RMSE). Although RMSE has a good physical and theoretical basis, it is often found to correlate very poorly with subjective ratings. A demonstration of this last fact can be found in Garcia et al. (2001) so as in some examples shown in this paper.

In order to avoid these drawbacks, some distinctness measures incorporating perceptual features have been developed. Some of these measures are based on single-channel models (Mannos and Sakrison, 1974; Saghri et al., 1989; Ahumada and Beard, 1998) and others on multichannel models (Rodriguez-Sanchez et al., 1999; Fdez-Vidal et al., 2000a; Garcia et al., 2001).

Recent studies (Rohaly et al., 1997; Ahumada and Beard, 1998) have shown that measures based on single-channel models achieve good performance in target distinctness. These measures have the advantage that are computationally less expensive in comparison with other measures based on more complex models.

Here, the optimal definition of a distinctness measure based on a single-channel model is estimated maximizing the degree of correlation with the visual target distinctness measured by human observers. We show that the derived optimal measure incorporates simple adaptive filtering, contrast masking and an error summation calculated only over a subset of points of the image (named interest points).

The approach is as follows. First, a psychophysical experiment is performed in which observers estimate the visual distinctness of targets in a database (Section 2). This subjective ranking induced by the psychophysical experiment is adopted as the reference rank order. Second, a computational distinctness measure is derived involving several factors, and applied to quantify the visual distinctness of the targets (Section 3). Several experiments are then performed to investigate the relation between the computational measure output and the psychophysical results (Section 4). Finally, the main conclusions are summarized in Section 5.

Section snippets

Psychophysical target distinctness

The images used in the psychophysical experiment are slides made during the distributed interactive simulation, search and target acquisition fidelity (DISSTAF) field test (Toet et al., 2001). These slides depict 44 different scenes. Each scene represents a military vehicle in a complex background (Fig. 1 shows some examples of targets). The visibility of the targets varies throughout the entire stimulus set. This is mainly due to variations in the structure of the local background, the viewing

Computational target distinctness

The computational measure predicts the target distinctness by measuring the difference between the target-and-background and the background-with-no-target scenes. Let t(x,y) be the image containing the target and e(x,y) the image without target.

The measure may be described in terms of three different stages: (1) first, a non-linear function is applied to the luminance values, (2) then, both images (t and e) are analyzed by a simple filter, and (3) finally, the target distinctness measure is

Experimental results

The images used in the computational experiments are those used in the psychophysical experiment. They are subsampled to 256×256 pixels. For each scene containing a target, a corresponding empty scene is created. The empty scene is everywhere equal to the target scene, except at the location of the target, where the target support is filled with the local background. This replacement is done by hand, using the rubber stamp tool in Photoshop 3.05. The result is judged by eye and is accepted if

Conclusions

The main conclusion drawn in this paper is that the derived computational measure UA1,C induces a visual target distinctness rank ordering that agrees with human visual perception for a set of complex natural scenes. Hence, a perceptual measure constructed with an only frequency channel to predict human performance in target detection must incorporate the following basic characteristics: (i) an adaptive CSF which depends on image distance; (ii) an estimation of image contrast to capture some of

Acknowledgements

This research was sponsored by the Dirección General de Enseñanza Superior (DGES) under grant PB98-1374.

References (18)

  • X.R. Fdez-Vidal et al.

    Integral opponent-colors features for computing visual target distinctness

    Pattern Recognition

    (2000)
  • A. Rohaly et al.

    Object detection in natural backgrounds predicted by discrimination performance and models

    Vision Res.

    (1997)
  • Ahumada, A.J., 1996. Simplified vision models for image quality assessment. In: SID International Digest of Technical...
  • Ahumada, A.J., Beard, B.L., 1998. A simple vision model for inhomogeneous image quality assessment. In: SID...
  • Barten, P., 1993. Spatio-temporal model for the contrast sensitivity of the human eye and its temporal aspects, In:...
  • J. Canny

    A computational approach to edge detection

    IEEE Trans. Pattern Anal. Machine Intell.

    (1986)
  • S. Daly

    The visible difference predictor: an algorithm for the assessment of image fidelity

  • B. Efron et al.

    An introduction to the bootstrap

    (1993)
  • X.R. Fdez-Vidal et al.

    Computing visual target distinctness through selective filtering, statistical features and visual patterns

    Opt. Eng.

    (2000)
There are more references available in the full text version of this article.

Cited by (0)

View full text