Two completely localized algorithms for deblurring shift-variant defocused images are presented. The algorithms exploit
limited support domain of 2D shift-variant point spread functions (PSFs) to localize the deblurring process. Focused
image at each pixel is modeled by a truncated Taylor-series polynomial and a localized equation is obtained which
expresses the blurred image as a function of the focused image and its derivatives. This localized equation forms the
basis of the two algorithms. The first algorithm iteratively improves the estimated focused image by directly evaluating
the localized equation for a given blurred image. The second algorithm uses the localized equation in a gradient descent
method to improve the focused image estimate. The algorithms use spatial derivatives of the estimate and hence exploit
smoothness to reduce computation. However, no assumptions about the blurring PSFs such as circular symmetry or
separability are required for computational efficiency. Due to complete localization, the algorithms are fully parallel, that
is, focused image estimates at each pixel can be computed independently. Performance of the algorithms is compared
quantitatively with other shift-variant image restoration techniques, both for computational efficiency and for robustness
against noise. The new algorithms are found to be faster and do not produce any blocking artifacts that are present in
sectioning methods for image restoration. Further, the algorithms are stable and work satisfactorily even in the presence
of large blur. Simulation results of the algorithms are presented for both Cylindrical and Gaussian PSFs. The
performance of the algorithms on real data is discussed.
KEYWORDS: 3D image processing, Point spread functions, Video, Cameras, 3D modeling, Motion models, Optical spheres, Computer simulations, Spherical lenses, Monte Carlo methods
A computationally efficient algorithm is described for generating shift-variant defocus and motion blur effects
for animation video. This algorithm precisely models rigid body motion of 3-D objects which includes arbitrary
translational and rotational motion. Camera parameters such as aperture diameter, focal length, and the location
of image detector are used to calculate the blur circle radius of point spread functions (PSFs) modeled by Gaussian
and Cylindrical functions. In addition, a novel and simple method similar to image inpainting is described for
filling missing pixels that arise due to object motion, round-off errors, interpolation or changes in magnification.
Performance of the algorithms are demonstrated on a set of 3D shapes such as sphere, cylinder, cone, etc. The
software tool developed in this research is also useful in computer vision and image processing research. It can
be used for simulating test data with known ground truth in the testing and evaluation of depth-from-defocus
and image/video de-blurring algorithms.
A localized and efficient iterative approach is presented for deblurring images that are highly defocused with
arbitrary shift-variant point spread functions (PSF). This approach extends a recently proposed local technique,
the RT technique, which works only for medium levels of blur to work for significantly higher levels of blur.
The RT technique is used to localize the blurring kernel at each pixel, and a region around the pixel with size
comparable to the size of the blurring kernel is divided into several smaller regions (intervals). The blurred image
in each interval is modeled separately by truncated Taylor-series polynomials. This step improves the accuracy
of the image model for low order truncated Taylor-series expansions. The blurred image value at each pixel is
expressed as the sum of multiple partial blur integrals with each integral term corresponding to one interval.
Then an expression is derived for the focused image value at a pixel in terms of the derivatives of the blurred
image in the central image region and solutions in the surrounding regions. This expression is solved iteratively
at each pixel in parallel to obtain a focused image. The starting solution is assumed to be either zero or the
blurred image itself. It is found that this new technique can effectively invert large blurs for which the original
RT method failed. In our experiments the truncated Taylor series expansion was limited to third order. Theory
and algorithms as well as experimental results on both simulation and real data are presented.
Image deblurring is an important preprocessing step in the inspection and measurement applications of machine vision
systems. A computational algorithm and analysis are presented for a new approach to one-dimensional shift-variant
image deblurring. The new approach is based on a new mathematical transform that restates the traditional shift-variant
image blurring model in a completely local but exactly equivalent form. The new approach is computationally noniterative,
efficient, and permits very fine-grain parallel implementation. The theory of the new approach for onedimensional
shift-variant deblurring is presented. Further, its advantages in comparison with related approaches, and
experimental results are presented.
A novel stereo camera architecture has been proposed by some researchers recently. It consists of a single digital
camera and a mirror adaptor attached in front of the camera lens. The adaptor functions like a pair of periscopes
which split the incoming light to form two stereo images on the left and right half of the image sensor. This novel
architecture has many advantages in terms of cost, compactness, and accuracy, relative to a conventional stereo
camera system with two separate cameras. However, straightforward extension of the traditional calibration
techniques were found to be inaccurate and ineffective. Therefore we present a new technique which fully
exploits the physical constraint that the stereo image pair have the same intrinsic camera parameters such as
focal length, principle point and pixel size. Our method involves taking one image of a calibration object and
estimating one set of intrinsic parameters and two sets of extrinsic parameters corresponding to the mirror
adaptor simultaneously. The method also includes lens distortion correction to improve the calibration accuracy.
Experimental results on a real camera system are presented to demonstrate that the new calibration technique
is accurate and robust.
A new passive ranging technique named Robust Depth-from-Defocus (RDFD) is presented for autofocusing in digital
cameras. It is adapted to work in the presence of image shift and scale change caused by camera/hand/object motion.
RDFD is similar to spatial-domain Depth-from-Defocus (DFD) techniques in terms of computational efficiency, but it
does not require pixel correspondence between two images captured with different defocus levels. It requires
approximate correspondence between image regions in different image frames as in the case of Depth-from-Focus (DFF)
techniques. Theory and computational algorithm are presented for two different variations of RDFD. Experimental
results are presented to show that RDFD is robust against image shifts and useful in practical applications. RDFD also
provides insight into the close relation between DFF and DFD techniques.
Depth From Defocus (DFD) is a depth recovery method that needs only
two defocused images recorded with different camera settings. In
practice, this technique is found to have good accuracy for cameras
operating in normal mode. In this paper, we present new
algorithms to extend the DFD method to cameras working in
macro mode used for very close objects in a distance range of
5 cm to 20 cm. We adopted a new lens position setting suitable for
macro mode to avoid serious blurring. We also developed a new
calibration algorithm to normalize magnification of images captured
with different lens positions. In some range intervals with high
error sensitivity, we used an additional image to reduce the error
caused by drastic change of lens settings. After finding the object
depth, we used the corresponding blur parameter for computing the
focused image through image restoration, which is termed as
"soft-focusing". Experimental results on a high-end digital camera
show that the new algorithms significantly improve the accuracy of
DFD in the macro mode. In terms of focusing accuracy, the RMS error
is about 15 lens steps out of 1500 steps, which is around 1%.
A new spatial-domain Blur Equalization Technique (BET) is presented. BET is based on Depth-from-Defocus
(DFD) technique. It relies on equalizing the blur or defocus of two different images recorded with different
camera parameters. Also, BET facilitates modeling of images locally by higher order polynomials with lower
series truncation errors. The accuracy of BET is further enhanced by discarding pixels with low Signal-to-Noise
ratio by thresholding image Laplacians, and relying more on sharper of the two blurred images in estimating
the blur parameters. BET is found to be superior to some of the best comparable DFD techniques in a large
number of both simulation and actual experiments. Actual experiments used a large variety of objects including
very low contrast digital camera test charts located at many different distances. In autofocusing experiments,
BET gave an RMS error of 1.2% in lens position.
Real-time and accurate autofocusing of stationary and moving objects is an important problem in modern digital cameras. Depth From Defocus (DFD) is a technique for autofocusing that needs only two or three images recorded with different camera parameters. In practice, there exist many factors that affect the performance of DFD algorithms, such as nonlinear sensor response, lens vignetting, and magnification variation. In this paper, we present calibration methods and algorithms for these three factors. Their correctness and effects on the performance of DFD have been investigated with experiments.
In this paper, several binary mask based Depth From Defocus (DFD) algorithms are proposed to improve autofocusing performance and robustness. A binary mask is defined by thresholding image Laplacian to remove unreliable points with low Signal-to-Noise Ratio (SNR). Three different DFD schemes-- with/without spatial integration and with/without squaring-- are investigated and evaluated, both through simulation and actual experiments. The actual experiments use a large variety of objects including very low contrast Ogata test charts. Experimental results show that autofocusing RMS step error is less than 2.6 lens steps, which corresponds to 1.73%. Although our discussion in this paper is mainly focused on a spatial domain method STM1, this technique should be of general value for different approaches such as STM2 and other spatial domain based algorithms.
A new technique is proposed for calibrating a 3D modeling system with variable zoom based on multi-view stereo image analysis. The 3D modeling system uses a stereo camera with variable zoom setting and a turntable for rotating an object. Given an object whose complete 3D model (mesh and texture-map) needs to be generated, the object is placed on the turntable and stereo images of the object are captured from multiple views by rotating the turntable. Partial 3D models generated from different views are integrated to obtain a complete 3D model of the object. Changing the zoom to accommodate objects of different sizes and at different distances from the stereo camera changes several internal camera parameters such as focal length and image center. Also, the parameters of the rotation axis of the turntable changes. We present camera calibration techniques for estimating the camera parameters and the rotation axis for different zoom settings. The Perspective Projection Matrices (PPM) of the cameras are calibrated at a selected set of zoom settings. The PPM is decomposed into intrinsic parameters, orientation angles, and translation vectors. Camera parameters at an arbitrary intermediate zoom setting are estimated from the nearest calibrated zoom positions through interpolation. A performance evaluation of this technique is presented with experimental results. We also present a refinement technique for stereo rectification that improves partial shape recovery. And the rotation axis of multi-view at different zoom setting is estimated without further calibration. Complete 3D models obtained with our techniques are presented.
KEYWORDS: 3D modeling, Visual process modeling, 3D image processing, 3D acquisition, Cameras, 3D vision, Image registration, Volume rendering, Data modeling, Data acquisition
A desktop vision system is presented for complete 3-D model reconstruction. It is fast (3-D reconstruction in under 20 min), low cost (uses a commercially available digital camera and a rotation stage), and accurate (about 1 part in 500 in the working range). Partial 3-D shapes and texture information are acquired from multiple viewing directions using rotational stereo and shape-from-focus (SFF). The resulting range images are registered to a common coordinate system, and a surface representation is created for each range image. The resulting surfaces are integrated using an algorithm named region of construction. Unlike previous approaches, the region of construction algorithm directly exploits the structure of the raw range images. The algorithm determines regions in the range images corresponding to nonredundant surfaces that can be stitched along the boundaries to construct a complete 3-D surface model. The algorithm is computationally efficient and less sensitive to registration error. It also has the ability to construct complete 3-D models of complex objects with holes. A textured 3-D model is obtained by mapping texture information onto the complete surface model representing the 3-D shape. Experimental results for several real objects are presented.
A new technique is introduced for registration and integration of multiple partial 3D models of an object. The technique exploits the epipolar constraint for the multiple- view geometry. Partial 3D shapes of an object from multiple viewing directions are obtained using a digital vision system based on parallel-axis stereo. The vision system is calibrated to obtain an initial transformation matrix for both the stereo imaging geometry and the multiple-view geometry. A multi-resolution stereo matching approach is used for partial 3D shape recovery. The partial 3D shapes are registered approximately using the initial transformation matrix. The initial transformation matrix is then refined by iteratively minimizing the registration error. At this step, a modified Iterative Closest Point (ICP) algorithm is used for matching corresponding points in two different views. A given point in one view is projected to another view using the transformation matrix, and a search is made for a closest point in the other view that lies on the epipolar line. A similar idea is used during partial model integration step to obtain improved results. Partial models are represented as linked lists of segments and integrated segment by segment. Experimental results are presented to show the effectiveness of the new technique.
KEYWORDS: 3D modeling, 3D image processing, Volume rendering, Cameras, Visual process modeling, 3D metrology, Image registration, 3D acquisition, Imaging systems, Reconstruction algorithms
New algorithms are presented for automatically acquiring the complete 3D model of single and multiple objects using rotational stereo. The object is placed on a rotation stage. Stereo images for several viewing directions are taken by rotating the object by known angles. Partial 3D shapes and the corresponding texture maps are obtained using rotational stereo and shape from focus. First, for each view, shape from focus is used to obtain a rough 3D shape and the corresponding focused image. Then, the rough 3D shape and focused images are used in rotational stereo to obtain a more accurate measurement of 3D shape. The rotation axis is calibrated using three fixed points on a planar object and refined during surface integration. The complete 3D model is reconstructed by integrating partial 3D shapes and the corresponding texture maps of the object from multiple views. New algorithms for range image registration, surface integration and texture mapping are presented. Our method can generate 3D models very fast and preserve the texture of objects. A new prototype vision system named Stonybrook VIsion System 2 (SVIS-2) has been built and used in the experiments. In the experiments, 4 viewing directions at 90-degree intervals are used. SVIS-2 can acquire the 3D model of objects within a 250 mm x 250 mm x 250 mm cubic workspace placed about 750 mm from the camera. Both computational algorithms and experimental results on several objects are presented.
KEYWORDS: 3D modeling, 3D image processing, Image analysis, Cameras, 3D acquisition, Data modeling, Volume rendering, Imaging systems, Data acquisition, 3D metrology
We present a digital vision system for acquiring the complete 3D model of an object from multiple views. The system uses image focus analysis to obtain a rough 3D shape of each view of an object and also the corresponding focused image or texture map. The rough 3D shape is used in a rotational stereo algorithm to obtain a more accurate measurement of 3D shape. The rotational stereo involves rotating the object by a small angle to obtain stereo images. It offers some important advantages compared to conventional stereo. A single camera is used instead of two, the stereo matching is easier as the field of view remains the same for the camera (but the object is rotated), and camera calibration is easier since a single stationary camera is used. The 3D shape and the corresponding texture map are measured for 4 views of the object at 90 degree angular intervals. These partial shapes and texture maps are integrated to obtain a complete 360 degree model of the object. The theory and algorithms underlying rotational- stereo and integration of partial 3D models are presented. The system can acquire the 3D model (which includes the 3D shape and the corresponding image texture) of a simple object within a 300mm x 300mm x 300mm volume placed about 600 mm from the camera. The complete model is displayed using a 3D graphics rendering software (Apple’s QuickDraw 3D Viewer). Both computational algorithms and experimental results on several objects are presented.
KEYWORDS: 3D modeling, Cameras, Visual process modeling, 3D image processing, Image analysis, 3D acquisition, 3D vision, Image restoration, Imaging systems, Computing systems
A digital vision system and the computational algorithms used by the system for three-dimensional (3D) model acquisition are described. The system is named Stonybrook VIsion System (SVIS). The system can acquire the 3D model (which includes the 3D shape and the corresponding image texture) of a simple object within a 300 mm X 300 mm X 300 mm volume placed about 600 mm from the system. SVIS integrates Image Focus Analysis (IFA) and Stereo Image Analysis (SIA) techniques for 3D shape and image texture recovery. First, 4 to 8 partial 3D models of the object are obtained from 4 to 8 views of the object. The partial models are then integrated to obtain a complete model of the object. The complete model is displayed using a 3D graphics rendering software (Apple's QuickDraw). Experimental results on several objects are presented.
A 3D vision system named SVIS is developed for 3D shape measurement that integrates three methods: (i) multiple- baseline, multiple-resolution Stereo Image Analysis (SIA) that uses colore image data, (ii) Image Defocus Analysis (IDA), and (iii) Image Focus Analysis (IFA). IDA and IFA are less accurate than stereo but they do not suffer from the correspondence problem associated with stereo. A rough 3D shape is first obtained using IDA and then IFA is used to obtain an improved estimate. The result is then used in SIA to solve the correspondence problem and obtain an accurate measurement of 3D shape. SIA is implemented using color images recorded at multiple-baselines. Color images provide more information than monochrome images for stereo matching. Therefore matching errors are reduced and accuracy of 3D shape is improved. Further improvements are obtained through multiple-baseline stereo analysis. First short baseline images are analyzed to obtain an initial estimate of 3D shape. In this step, stereo matching errors are low and computation is fast since a shorter baseline result in lower disparities. The initial estimate of 3D shape is used to match longer baseline stereo images. This yields more accurate estimation of 3D shape. The stereo matching step is implemented using a multiple-resolution matching approach to reduce computation. First lower resolution images are matched and the result are used in matching higher resolution images. This paper presented the algorithms and the experimental result of 3D shape measurements on SVIS for several objects. These results suggest a practical vision system for 3D shape measurement.
The theory of Unified Focus and Defocus Analysis (UFDA) was presented by us earlier and it was extended to use both classical optimization technique and regularization approach for 3D scene recovery. In this paper we present a computational algorithm for UFDA which uses variable number of images in an optimal fashion. UFDA is based on modeling the sensing of defocused images in a camera system. This approach unifies Image Focus Analysis (IFA) and Image Defocus Analysis (IDA), which form two extremes in a range of possible methods useful in 3D shape and focused image recovery. The proposed computational algorithm consists of two main steps. In the first step, an initial solution is obtained by a combination of IFA, IDA, and interpolation. In the second step, the initial solution is refined by minimizing the error between the observed image data and the image data estimated using a given solution and the image formation model. A classical gradient descent or a regularization technique is used for error minimization. Our experiments indicate that the most difficult part of the algorithm is to obtain a reasonable solution for the focused image when only a few image frames are available. We employ several methods to address this part of the problem. The algorithm has been implemented and experimental results are presented.
KEYWORDS: 3D image processing, Image analysis, Image restoration, Computing systems, Cameras, Image processing, Data modeling, Data analysis, Point spread functions, 3D modeling
A unified approach to image focus and defocus analysis (UFDA) was proposed recently for three-dimensional shape and focused image recovery of objects. One version of this approach which yields very accurate results is highly computationally intensive. In this paper we present a parallel implementation of this version of UFDA on the Parallel Virtual Machine (PVM). One of the most computationally intensive parts of the UFDA approach is the estimation of image data that would be recorded by a camera for a given solution for 3D shape and focused image. This computational step has to be repeated once during each iteration of the optimization algorithm. Therefore this step has been sped up by using the Parallel Virtual Machine (PVM). PVM is a software package that allows a heterogeneous network of parallel and serial computers to appear as a single concurrent computational resource. In our experimental environment PVM is installed on four UNIX workstations communicating over Ethernet to exploit parallel processing capability. Experimental results show that the communication over-head in this case is relatively low. An average of 1.92 speedup is attained by the parallel UFDA algorithm running on 2 PVM connected computers compared to the execution time of sequential processing. By applying the UFDA algorithm on 4 PVM connected machines an average of 3.44 speedup is reached. This demonstrates a practical application of PVM to 3D machine vision.
Image defocus analysis (IDA), image focus analysis (IFA), and stereo image analysis (SIA), are integrated for recovering the three-dimensional (3D) shape of objects. Integrating IDA, IFA, and SIA, has important advantages because IDA and IFA are less accurate than stereo but they do not suffer from the correspondence problem associated with stereo. Therefore, a rough 3D shape is first obtained using IDA and IFA without encountering the correspondence problem. The amount of computation and accuracy at this stage is optimized by using IDA first and then the IFA. Accuracy is further improved by projecting a high contrast pattern on to the object of interest. The rough shape thus obtained is used in a stereo matching algorithm to solve the correspondence problem efficiently. The amount of computation in matching is reduced since the search for correspondence is done in a narrow image region determined by the approximate shape. Also, the knowledge of approximate shape improves the matching accuracy by minimizing false matches due to occlusion. The method for integrating IDA, IFA, and SIA is presented in detail. The method has been implemented on a vision system named SVIS. Experimental results of the method are presented.
KEYWORDS: 3D image processing, Image analysis, 3D modeling, Shape analysis, Point spread functions, Image restoration, 3D image reconstruction, 3D acquisition, Data acquisition, Data modeling
The reconstruction of three-dimensional (3D) information from defocused image data is formulated as an inverse-problem that is solved through a regularization technique. The technique is based on modeling the sensing of defocused images in a camera system using a three-dimensional (3D) point spread function (PSF). Many images are acquired at different levels of defocus. The difference (mean-square error) between this acquired image data and the estimated image data corresponding to an initial solution for 3D shape is minimized. The initial solution for 3D shape is obtained from a focus and defocus analysis approach. A regularization approach that uses a smoothness constraint is proposed to improve this initial solution iteratively. The performance of this approach is compared with two other approaches: (1) gradient descent based on planar surface patch approximation, and (2) a local error minimization based on a limited search. We exploit some constraints such as the positivity of image brightness unique to this problem in the optimization procedure. Our experiments show that the regularization approach performs better than the other two and that high accuracy is attainable with relatively moderate computation. Experimental results are demonstrated for geometric optics model of 3D PSF on simulated image data.
Depth-from-Defocus using the Spatial-Domain Convolution/Deconvolution Transform Method (STM) is a useful technique for 3D vision. STM involves simple local operations in the spatial domain on only two images recorded with different camera parameters (e.g. by changing lens position or changing aperture diameter). In this paper we provide a theoretical treatment of the noise sensitivity analysis of STM and verify the theoretical results with experiments. This fills an important gap in the current research literature wherein the noise sensitivity analysis of STM is limited to experimental observations. Given the image and noise characteristics, here we derive an expression for the Root Mean Square (RMS) error in lens position for focusing an object. This RMS error is useful in estimating the uncertainty in depth obtained by STM. We present the results of computer simulation experiments for different noise levels. The experiments validate the theoretical results.
KEYWORDS: Image analysis, Cameras, 3D image reconstruction, 3D image processing, Point spread functions, 3D modeling, Data analysis, Image sensors, Sensors, Analytical research
A new approach is proposed for highly accurate reconstruction of 3D shape and focused image of an object from a sequence of noisy defocused images. This approach unifies the two approaches: image focus analysis and image defocus analysis, which have been treated separately in the research literature so far. In the new unified approach, high accuracy is attained at the cost of increased data acquisition and computation. This approach is based on modeling the sensing of defocused images in a camera system. A number of images are acquired at different levels of defocus. The resulting data is treated as a function sampled in the 3D space where x and y are the image spatial coordinates and d is a parameter representing the level off defocus. The concept of a '3D point spread function' in the space is introduced. The problem of 3D shape and focused image reconstruction is formulated as an optimization problem where the difference between the observed image data and the estimated image data is minimized. The estimated image data is obtained from the image sensing model and the current best known solutions to the 3D shape and focused image. An initial estimation to the solutions is obtained through traditional shape-from-focus methods. This solution is improved iteratively by a gradient descent approach. This approach reduces the errors in shape and focused image introduced by the image-overlap problem and the non- smoothness of the object's 3D shape. Experimental results are presented to show that the new method yields improved accuracy.
Image focus analysis is an important technique for passive autofocusing and 3D shape measurement. Electronic noise in digital images introduces errors in this techniques. It is therefore important to derive robust focus measures that minimize error. In our earlier research, we have developed a method for noise sensitivity analysis of focus measures. In this paper we derive explicit expressions for the root-mean square (RMS) error in autofocusing based on image focus analysis. This is motivated by the autofocusing uncertainty measure (AUM) defined earlier by us as a metric for comparing the noise sensitivity of different focus measures in autofocusing and 3D shape-from-focus. The RMS error we derive by us has the same advantage as AUM in that it can be computed in only one trial of autofocusing. We validate our theory on RMS error and AUM through experiments. It is shown that the theoretically estimated and experimentally measured values of the standard deviation of a set of focus measures are in agreement. Our results are based on a theoretical noise sensitivity analysis of focus measures, and they show that for a given camera the optimally accurate focus measure may change from one object to the other depending on their focused images.
The optimally accurate focus measure for a noisy camera in passive search based autofocusing and depth-from-focus applications depends not only on the camera characteristics but also the image of the object being focused or ranged. In this paper a new metric named autofocusing uncertainty measure (AUM) is defined which is useful in selecting the most accurate focus measure from a given set of focus measures. AUM is a metric for comparing the noise sensitivity of different focus measures. It is similar to the traditional root-mean-square (RMS) error, but, while RMS error cannot be computed in practical applications, AUM can be computed easily. AUM is based on a theoretical noise sensitivity analysis of focus measures. In comparison, all known work on comparing the noise sensitivity of focus measures have been a combination of subjective judgement and experimental observations. For a given camera, the optimally accurate focus measure may change from one object to the other depending on their focused images. Therefore selecting the optimal focus measure from a given set involves computing all focus measures in the set. However, if computation needs to be minimized, then it is argued that energy of the Laplacian of the image is a good focus measure and is recommended for use in practical applications.
A Depth-from-Defocus method named STM was presented recently for stationary objects. Here we extend STM for continuous focusing of moving objects. The method is named Continuous STM or CSTM. Focusing is done by moving the lens with respect to the image detector. Two variations of CSTM - CSTM1 and CSTM2 - are presented. CSTM1 is a straight forward extension of STM described in. It involves calibration of the camera for a number (about 6 in our implementation) of discrete lens positions. In CSTM2 the camera is calibrated only for one lens position. The calibration data corresponding to other lens positions are obtained by transforming the data of the one lens position for which the camera is calibrated. In the experimental results presented here, the focusing error in lens position was about 2.25% for CSTM1 and about 3% for CSTM2.
A method is presented for continuous focusing of moving objects. It is based on DFD1F which is a method for determining distance of stationary objects using image defocus information. The proposed method requires recording of two images of a moving object with different degrees of blur. The change in blur is caused by varying camera parameters such as lens position, focal length, and aperture diameter of the recording camera. The two images must be recorded simultaneously in a short time period. A new camera structure is proposed for such recording of the images. In the proposed method, the requirement of a large memory space has been avoided for storing the MTF data of the camera's optical system. This is achieved by using a parameterization scheme for the MTF data. The method has been implemented on an actual camera system. Experimental results on this system indicate that the method yields an RMS error in focusing of about 4.3 percent in lens position. The image blur caused by a focusing error of this magnitude is barely noticeable by humans. Therefore, in addition to machine vision, the method has practical applications in video cameras such as camcorders.
Verification of computer vision theories is facilitated by the development and implementation of computer simulation systems. Computer simulation avoids the necessity of building actual camera systems; they are fast, flexible, and can be easily duplicated for use by others. In our previous work, we proposed a useful computational model to explore the image sensing process. This model decouples the photometric information and the geometric information of objects in the scene. In this paper, we further extend the proposed image sensing model to simulate the image formation of moving objects (motion) and stereo vision system. The simulation algorithms for curved objects, moving objects, and stereo imaging are presented. Based on the proposed model and algorithms, a computer simulation system called Active Vision Simulator (AVS) has been implemented. AVS can be used to simulate image formation process in a monocular (MONO mode) or binocular (STEREO mode) camera system to synthesize the images. It is useful for research on image restoration, motion analysis, depth from defocus, and algorithms for solving the correspondence problem in stereo vision. The implementation of AVS is efficient, modular, extensible, and user-friendly.
We use the paraxial geometric optics model of image formation to derive a set of camera focusing techniques. These techniques do not require calibration of cameras but involve a search of the camera parameter space. The techniques are proved to be theoretically sound under weak assumptions. They include energy maximization of unfiltered, low-pass-filtered, high-pass-filtered, and bandpass-filtered images. It is shown that in the presence of high spatial frequencies, noise, and aliasing, focusing techniques based on bandpass filters perform well. The focusing techniques are implemented on a prototype camera system called the Stonybrook passive autofocusing and ranging camera system (SPARCS). The architecture of SPARCS is described briefly. The performance of the different techniques are compared experimentally. All techniques are found to perform well. The energy of low-pass-filtered image gradient, which has better overall characteristics, is recommended for practical applications.
A new shape-from-focus method is described which is based on a new concept named Focused Image Surface (FIS). FIS of an object is defined as the surface formed by the set of points at which the object points are focused by a camera lens. According to paraxial-geometric optics, there is a one-to-one correspondence between the shape of an object and the shape of its FIS. Therefore, the problem of shape recovery can be posed as the problem of determining the shape of the FIS. From the shape of FIS the shape of the object is easily obtained. In this paper the shape of the FIS is determined by searching for a shape which maximizes a focus measure. In contrast with previous literature where the focus measure is computed over the planar image detector of the camera, here the focus measure is computed over the FIS. This results in more accurate shape recovery than the traditional methods. Also, using FIS, a more accurate focused image can be reconstructed from a sequence of images than is possible with traditional methods. The new method has been implemented on an actual camera system, and the results of shape recovery and focused image reconstruction are presented.
A computer model is presented for the formation and sensing of defocused images in a typical CCD camera system. A computer simulation system named IDS has been developed based on the computer model. IDS takes as input the camera parameters and the scene parameters. It produces as output a digital image of the scene as sensed by the camera. IDS consists of a number of distinct modules each implementing one step in the computer model. The modules are independent and can be easily modified to enhance the simulation system. IDS is being used by our group for research on methods and systems for determining depth from defocused images. IDS is also being used for research on image restoration. It can be easily extended and used as a research tool in other areas of machine vision. IDS is a machine independent and hence portable. It provides a friendly user interface which gives the user full access and control to the parameters and intermediate results.
This paper describes the application of a new Spatial-Domain Convolution/Deconvolution transform (S transform) for determining distance of objects and rapid autofocusing of camera systems using image defocus. The method of determining distance, named STM, involves simple local operations on only a few (about 2 to 4) images and it can be easily implemented in parallel. STM has been implemented on an actual camera system named SPARCS. Experiments on the performance of STM and their results on real-world objects are presented. The results indicate that STM is useful in practical applications. The utility of the method is demonstrated for rapid autofocusing of electronic cameras. STM is computationally more efficient than other methods, but for our camera system, it is somewhat less robust in the presence of noise than a Fourier transform based approach. STM is a useful technique in many applications such as rapid autofocusing.
We use the paraxial geometric optics model of image formation to derive a set of camera focusing techniques. These techniques do not require calibration of cameras but involve a search of the camera parameter space. The techniques are proved to be theoretically sound. They include energy maximization of unfiltered, low-pass filtered, high-pass filtered, and band-pass filtered images. It is shown that in the presence of high spatial frequencies, noise, and aliasing, focusing techniques based on band-pass filters perform well. The focusing techniques are implemented on a prototype camera system named SPARCS. The architecture of SPARCS is described briefly. The performance of the different techniques are compared experimentally. All techniques are found to perform well. One of them -- the energy of low pass filtered image gradient -- which has better overall characteristics is recommended for practical applications.
A mathematical model for a typical CCD camera system used in machine vision applications is presented. This model is useful in research and development of machine vision systems and in the computer simulation of camera systems. The model has been developed with the intention of using it to investigate algorithms for recovering depth from image blur. However the model is general and can be used to address other problems in machine vision. The model is based on a precise definition of input to the camera system. This definition decouple3 the photometric properties of a scene from the geometric properties of the scene in the input to the camera system. An ordered sequence of about 20 operations are defined which transform the camera system''s input to its output i. e. digital image data. Each operation in the sequence usually defines the effect of one component of the camera system on the input. This model underscores the complexity of the actual imaging process which is routinely underestimated and oversimplified in machine vision research.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.