Log-polar mapping template design: From task-level requirements to geometry parameters☆
Introduction
The computational interest of the log-polar image model [5], which has a biological inspiration [32], has been explored over about two decades by the pattern recognition [44] and the active vision [7] communities. As a practical issue in computer vision, log-polar images can be obtained through camera research prototypes [14], [45], [27], by software simulations [3], [25], [36], or with programmable hardware cards [16], [12]. While log-polar images are a direct output of the visual sensor in the first case, cartesian images are required as input to the log-polar mapping procedures of the two other cases.
The design and fabrication of log-polar sensors has to deal with a number of technical and technological difficulties, but significant improvements have been achieved, including a color log-polar CMOS sensor [31]. Unlike the many obstacles and high economic cost of this option, obtaining log-polar images from cartesian images by means of software is a very flexible and cheap alternative. Unfortunately, this simplicity probably helps to explain why the proper choice of the mapping parameters has been mostly ignored in the past.
However, several reasons suggest the significance of devising adequate criteria when designing a log-polar mapping template. First of all, it is important to bear in mind that the particular log-polar geometry being used may affect the performance of some visual tasks using these images. Second, users of procedures to obtain log-polar would benefit from the availability of guidelines to get good mapping templates. Finally, designers interested in improving their log-polar sensors might follow some existing set of design criteria.
The impact of the log-polar mapping parameters on the performance of a vergence control task was analyzed in [2]. The relevance of an appropriate parameter selection was also highlighted in [1]. Precision requirements of 3D measurements led to a choice of log-polar geometry parameters in [40]. Possibilities for foveae design were considered at length in [42], while the quality of log-polar templates were quantified in [30], [31]. This summarizes the few works related to the problem of selecting the log-polar sensor parameters. All of them tend to focus on a particular visual problem or lack a more general design perspective accounting for the variety of visual tasks which are possible. In contrast, this article studies a set of general log-polar mapping template design criteria, as well as a mechanism to find the mapping parameters meeting those criteria. A general framework aimed at reducing the knowledge gap between task requirements and low-level mapping parameters is also suggested.
In this article, the problem of sensor design is formulated as an optimization problem. More specifically, a multi-objective genetic algorithm is used. A rich literature exists on parameter optimization in the context of computer vision systems. The scope of these systems may range from the design of optimal image filters [29], to finding the best parameters of a motion detection system [6]. An area of recent interest is that of discovering an optimal placement of a set of visual sensors [8], [23], [48], so that the total price can be minimized or the coverage can be maximized. The design of catadioptric sensors has also been considered lately [13], [20], [21]. The work in this article shares with these the fact that the proper configuration of a vision system is designed by means of some kind of optimization.
The work presented here is based on our preliminary efforts [37], and represents improvements in several aspects. First, the overall perspective of the work has been enhanced by including a proposal of a general design framework. Second, the design criteria have been better formalized, and the importance of having unit aspect-ratio receptive fields has been discussed more rigorously. Third, several examples illustrating the proposed approach are given. Finally, and most importantly, the optimization algorithm finding the sensor parameters from a set of design criteria is now much more general and effective. For the sake of clarity and completeness of this article, some background information from this previous paper [37], is part of Sections 1 Introduction, 3 Log-polar sensor: the low-level mapping parameters, 4 Design criteria, 5 From task requirements to design criteria.
The rest of this article is organized as follows. First, a general design framework is described (Section 2). Next, the log-polar model and its basic parameters are introduced (Section 3). Then, a set of general design criteria are discussed (Section 4), and examples of task requirements are given (Section 5). Subsequently, the multi-objective genetic algorithm in charge of finding the primitive parameters best fulfilling a set of given design criteria is presented (Section 6). Illustrative examples of sensor designs found following the described procedure are then provided (Section 7). Finally, concluding remarks are given (Section 8).
Section snippets
Overview of the approach
Generally, the best configuration for a given sensor1 depends on the particular task in which the sensor is to be used. Therefore, since sensor design is task-dependent, some procedure is required to translate task requirements into low-level sensor parameters. If this procedure is not automatic, human intervention is needed,
Log-polar sensor: the low-level mapping parameters
Although there are several possible log-polar image models available [7], the central blind-spot model was chosen because of its implementation simplicity and the interesting edge invariance property2 [44].
Design criteria
From the basic parameters involved in the log-polar transform, we define other, derived, parameters which are a quantification of some properties of the log-polar mapping template. A particular value for each of these derived parameters represents a design criterion. Our previous work [37] included these criteria, except an additional one, the minimum resolution (Section 4.5).
From task requirements to design criteria
Now we give examples of visual task requirements which could be required in practical, realistic scenarios. We also show how they can be translated into design criteria.
From design criteria to mapping parameters
Finally, to complete the picture, we explain how a set of design criteria can be translated into low-level, mapping parameters. First, it is explained how design criteria can be represented (Section 6.1). Then, the choice of genetic algorithms as the sensor-design expert is justified (Section 6.2). Next, it is discussed how the representation of design criteria can be enhanced by including relevance weights (Section 6.3), which should be handled by the expert. The genetic algorithm finding the
Examples
To illustrate the proposed system, four examples are detailed below. In each of them, task requirements for a particular visual task are described, then they are translated into design criteria, and the resulting sensor parameters, as found by the genetic algorithm, are provided.
Conclusions
Three facts motivated this work. First, in most works in the literature, the choice for the values of the parameters of the log-polar transform was found to be arbitrary or weakly justified. Second, the best configuration of the sensor is a matter of the visual task at hand. And third, users of a visual sensor cannot be assumed to have a deep knowledge of the sensor or its design parameters.
To overcome these problems, a three-level framework is proposed to separate task requirements from design
References (48)
- et al.
Visual behaviors for binocular tracking
Robotics and Autonomous Systems
(1998) Optimizing parameters of a motion detection system by means of a distributed genetic algorithm
Image and Vision Computing (IVC)
(2005)- et al.
A review of biologically motivated space-variant data reduction models for robotic vision
Computer Vision and Image Understanding (CVIU)
(1998) - et al.
Face detection: a survey
Computer Vision and Image Understanding (CVIU)
(2001) - et al.
A saliency-based search mechanism for overt and covert shifts of visual attention
Vision Research
(2000) A new log-polar mapping for space variant imaging. Application to face detection and tracking
Pattern Recognition
(1999)- et al.
Bias minimizing filter design for gradient-based image registration
Signal Processing: Image Communication
(2005) - et al.
Object-based visual attention for computer vision
Artificial Intelligence
(2003) - et al.
Dealing with 2D translation estimation in log-polar imagery
Image and Vision Computing (IVC)
(2003) - et al.
Similarity motion estimation and active tracking through spatial-domain projections on log-polar images
Computer Vision and Image Understanding (CVIU)
(2005)
Binocular tracking: integrating perception and control
IEEE Transactions on Robotics and Automation
Anthropomorphic visual sensors
Grid coverage for surveillance and target location in distributed sensor networks
IEEE Transactions on Computers
An updated survey of GA-based multiobjective optimization techniques
ACM Computing Surveys
Real-time binocular smooth pursuit
International Journal of Computer Vision
Evolutionary Computation: A Unified Approach
Stochastic local search for omnidirectional catadioptric stereovision design
A foveated retina-like sensor using CCD technology
Genetic Algorithms in Search, Optimization, and Machine Learning
Cited by (34)
Modeling and simulations of three-dimensional laser imaging based on space-variant structure
2016, Optics and Laser TechnologyAn integrated neuromimetic architecture for direct motion interpretation in the log-polar domain
2014, Computer Vision and Image UnderstandingCitation Excerpt :To obtain such an “interaction”, the relationships among the cortical descriptors and the Cartesian ones in the area of interest must be devised (see Section 3.5). Different models for transforming the Cartesian images into the log-polar domain have been proposed [55,56,51]. The blind-spot model is particularly interesting, for its properties of rotation and scale invariance, and for the simplicity of its implementation.
Modulation transfer function of spatially variant sampling retina-like sensor
2013, OptikCitation Excerpt :Retina-like sensors, which are shown in Fig. 1, mimic the distribution of photoreceptors in the human retina. They are fabricated and used in the fields of robot vision, image transmission, target tracking, real-time control, emergency communications, three-dimensional (3D) perception, omni-directional camera, egomotion detection and target detection based on task [2–8]. The output property of retina-like sensor is superior to that of traditional rectilinear sensor during high-speed forward motion imaging [9].
Design strategies for direct multi-scale and multi-orientation feature extraction in the log-polar domain
2012, Pattern Recognition LettersCitation Excerpt :The aspect-ratio of the log-polar pixel has to be around 1. It is worth noting that in (Traver and Pla, 2008) the authors state that a log-polar pixel with aspect ratio equals to 1 is necessary to correctly compute the gradient orientation. The analysis conducted in this paper shows that this rule can be generalized in order to efficiently use the local spatial operators to measure important elements of the visual signal (Adelson and Bergen, 1991; Granlund and Knutsson, 1995).
Robust Focus Measure Operator Using Adaptive Log-Polar Mapping for Three-Dimensional Shape Recovery
2015, Microscopy and MicroanalysisNear-optimal combination of disparity across a log-polar scaled visual field
2020, PLoS Computational Biology
- ☆
Acknowledgment to projects P1-1B2004-33 and P1-1A2004-02, funded by Fundació Caixa-Castelló Bancaixa, and to projects HP2005-0095 (Integrated Actions) and CSD2007-00018 (Consolider Ingenio 2010), both funded by the Spanish Ministerio de Educación y Ciencia.