2006 International Joint Conference on Neural Networks
Sheraton Vancouver Wall Centre Hotel, Vancouver, BC, Canada
July 16-21, 2006
Neural-Network-Based Photometric Stereo for 3D Surface
Reconstruction
Wen-Chang Cheng
Department of Information Network Technology
Hsiuping institute of Technology
11 Gungye Road, Dali City, Taiwan.
E-mail: wccheng@mail.hit.edu.tw
Abstract—This paper proposes a novel neural-network-based
photometric stereo approach for 3D surface reconstruction.
The neural network inputs are the pixel values of the 2D
images to be reconstructed. The normal vectors of the surface
can then be obtained from the weights of the neural network
after supervised learning, where the illuminant direction does
not have to be known in advance. Finally, the obtained normal
vectors are applied to enforce integrability when reconstructing
3D objects. The experimental results demonstrate that the
proposed neural-network-based photometric stereo approach
can be successfully applied to objects generally, and perform
3D surface reconstruction better than some existing
approaches.
Keywords: Lambertian model, shape from shading,
enforcing integrability, surface normal, neural network.
1. Introduction
Photometric stereo approach is able to estimate
local surface orientation by using several images of the
same surface taken from the same viewpoint but under
illuminations from different directions. It was first
introduced based on the Lambertian reflectance model
by R. J. Woodham[1]. It has received wide attention
and several efforts have been made to improve the
performance of recovery [2]-[23]. The main limitation
of classical photometric stereo approach is that the
light source positions must be accurately known and
this necessitates a fixed, calibrated lighting rig. Hence,
an improved photometric stereo method for estimating
the surface normal and the surface reflectance of
objects without a priori knowledge of the light source
direction or the light source intensity is proposed by
Hayakawa [19]. The method used the singular-value
decomposition (SVD) method to factorize image data
matrix of three different illuminations into surface
reflectance matrix and light source matrix based on the
Lambertian model. However, they still used one of the
two added constraints (i.e., at least 6 pixels in which
relative value of the surface reflection is constant or
1
0-7803-9490-9/06/$20.00/©2006 IEEE 404
Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.
known and at least 6 frames in which the relative value
of the light-source intensity is constant or known) for
finding the linear transformation between the surface
reflectance matrix and the light source matrix.
Belhumeur etc. [10] showed that a generalized
bas-relief transformation is a transformation of both
the surface shape and the surface albedo for an
arbitrary Lambertian surface. The set of images of an
object in fixed post but under all possible illumination
conditions is a convex cone (illumination cone) in the
space of images. When the surface reflectance can be
approximated as Lambertian, this illumination cone
can be constructed from a handful of images acquired
under variable lighting. They used as few as seven
images of a face seen in a fixed pose, but illumination
by point light sources at varying, unknown position, to
estimate its surface geometry and albedo map up to a
generalized bas-relief transformation. Despite they
announced their success under unknown light source
directions, the estimation of surface methods still need
to be assisted with some added constraints or more
images.
Multi-layer neural networks have also been
adopted to handle the photometric stereo problem
[20-22]. However, these approaches are still restricted
by the Lambertian model, which requires estimating
the direction of the light source. Obviously, this
restriction makes the algorithm impractical for many
applications in which illumination information is not
available. In this paper, we propose a novel
neural–network–based photometric stereo method to
solve the Lambertian model for 3D surface recognition.
A supervised learning algorithm is applied to tune up
the normal vectors of the surface for reconstruction and
light source vectors automatically based on image
intensities. The 3D surface can also be reconstructed
according to these normal vectors using existing
approaches such as enforcing integrability [24].
Additionally, the proposed approach considers the
albedo and the reflecting characteristic of each surface
point individually. According to the experimental
results presented in Section 6, the shape recovery
algorithm is the most robust for recovery of surfaces
with the variant albedo and complex reflecting
characteristic.
The remainder of this study is organized as
follows. Section 2 describes the Lambertian reflectance
model. The details of the neural-network- based
photometric stereo approach and its learning rule
derivations are presented in Sections 3 and 4. Section 5
introduces the 3D surface reconstruction from the
surface normal by enforcing integrability. Section 6
presents the results of experiments performed to
evaluate the performance of the proposed approach.
Conclusions are drawn in the last section.
2. The Lambertian reflectance model
Lambertian model which represents a surface
illuminated by a single point light source given by
Rd (n ( x, y ), α ( x, y )) = L α ( x, y ) n ( x, y ) T s , (1)
where α ( x, y ) denotes the surface’s albedo at
point ( x, y ) over the image plane domain; n( x, y)
denotes the surface normal on position ( x, y ) ; s ∈ R 3
denotes the direction of point light at infinity, and L
denotes the light strength. Assuming that the
Lambertian surface of a convex object is given by the
depth function z ( x, y ) , the surface normal n( x, y )
on position ( x, y ) can be represented as
n ( x, y ) =
[− z x ( x, y) − z y ( x, y ) 1 T ] ,
z x2 ( x, y) + z 2y ( x, y) + 1
where z x ( x, y ) and z y ( x, y ) denote the x- and y-
partial derivatives of z ( x, y ) , respectively. Therefore,
the 3D surface can also be reconstructed according to
these normal vectors using the enforcing integrability
approach (The details shows in the Section 5).
3. Our Proposed Method to Solve the Lambertian
Reflectance Model
Figure 1 shows the framework of the proposed
symmetric neural network which simulates the
Lambertian reflection model. The input/output pairs of
the network are arranged like a mirror in the center
layer, where the number of input nodes equals the
number of output nodes, making it a symmetric neural
network. The light source direction and the normal
Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.
vector from the input 2D images in the left side of the
symmetric neural network are separated and then
combined inversely to generate the reflectance map in
the right side of the network. The function of each
layer is discussed in detail below.
Figure 1. Framework of the symmetric neural network for
diffuse reflection model.
Assuming that an input image has m pixels in
total, then the symmetric neural network includes m
input variables. The 2D image is rearranged to form a
m ×1 column vector represented as
I = (I 1 , I 2 , ..., I m )T and fed into the symmetric neural
network. Through the symmetric neural network, the
reflectance map R = (R1 , R2 , ..., Rm )T can be
obtained in the output of the symmetric neural
network.
Each node also outputs an activation value as a
function of its net-input,
node-output = a (l ) (net-input) = a (l ) ( f ) , (3)
(2) where a (l ) (⋅) denotes the activation function, and the
superscript l denotes the layer number. The proposed
symmetric neural network contains six layers. The
functions of the nodes in each layer are described as
follows.
Layer 1: This layer gathers the intensity values of
the input images as the network inputs. Node I i
denotes the ith pixel of the 2D image, and m denotes
the number of total pixels of the image. That is,
Ii
fi = , i = 1, ..., m,
max ( I j ) (4)
1≤ j ≤ m
ai(1) = f i , i = 1, ..., m.
The following equation also uses this notation.
Layer 2: This layer adjusts the intensity of the
input 2D image with corresponding albedo value. Each
node in this layer, corresponding to one input variable,
2
405
divides the input intensity by the corresponding albedo
and transmits it to the next layer. That is,
Ii 1 j =1
fi = ⋅ , i = 1, ..., m,
max ( I j ) α i
1≤ j ≤m (5)
Iˆi = ai( 2) = f i , i = 1, ..., m.
The output of this layer is the adjusted intensity value
of the original normalized 2D image. The nodes of this
layer are labeled as Iˆ1 , Iˆ2 , ..., Iˆm . The term α i
denotes the ith albedo value corresponding to the ith
pixel of the 2D image, and 1 α i denotes the weight
between I and Iˆ .
i i
Layer 3: The purpose of layer 3 is to separate the
light source direction from the 2D image. The light
source directions of this layer is not normalized, and
are labeled as s′1 , s ′2 , and s′3 . The link weight in
layer 3 is represented as wij for the connection
between node i of layer 2 and node j of layer 3.
m
f j = ∑ Iˆi wij , i = 1, ..., m, j = 1, 2, 3,
i =1
s′j = a (j3) = f j, j = 1, 2, 3.
Layer 4: The nodes of this layer represent the
unit light source. Equation 6 is used to normalize the
non-normalized light source direction obtained in layer
3. These nodes in layer 4 are labeled as s1 , s 2 , and s3 ,
respectively, and the light source direction is
represented as s = (s1 , s 2 , s3 ) . The output of s j can
T
be calculated from:
1 image; Ri denotes the ith output of the neural
fj = , j = 1, 2, 3,
s1′ + s 2′ 2 + s 3′ 2
2
s ′j
s j = a (j4) = f j ⋅ s ′j =
s1′ 2 + s 2′ 2 + s3′ 2
(7)
Layer 5: This layer 5 combines the light source
direction s and normal vectors of the surface to
generate the reflectance map. The link weight
connecting node j of layer 4 and node k of layer 5 is
denoted as v jk , and represents the normal vectors of
the surface. That is, (v1k , v 2k , v 3k ) T denotes the
normal vector of the surface on the point k, where k =
1, …, m. The outputs of the nodes in this layer are
denoted as R̂k , and can be calculated as:
3
f k = ∑ s j v jk , k = 1, ..., m,
(8)
Rˆ k = a k(5) = fk , k = 1, ..., m.
Significantly, R̂k denotes the non-normalized
reflectance map, and therefore is normalized in Layer
6.
Layer 6: This layer transfers the non-normalized
reflectance map obtained in Layer 5 into the interval [0,
1]. These nodes, R1 , R2 , ..., Rm , denote the normalized
reflectance map, and their output can be calculated as:
f k = Rˆ k , k = 1, ..., m,
R k = a k( 6) =
( f k − min(Rˆ )) (9)
max (R ˆ ) − min (R
ˆ)
=
(Rˆ k − min(Rˆ )) , k = 1, ..., m,
max (R ˆ ) − min (R
ˆ)
where Rˆ = ( Rˆ1 , Rˆ 2 , ..., Rˆ m ) T , and the link weights
between layers 5 and 6 are unity.
Through the supervised learning algorithm
(6) derived in the following section, the normal surface
vectors can be obtained automatically.
4. Training Algorithm of the Proposed Model
Back-propagation learning is employed for
supervised training of the proposed model to minimize
the error function defined as
m
ET = ∑ (Ri − Di ) ,
2
(10)
i =1
where m denotes the number of total pixels of the 2D
network, and Di denotes the ith desired output equal
to the ith intensity of the original normalized 2D image.
For each 2D image, starting at the input nodes, a
, j = 1, 2, 3.
forward pass is used to calculate the activity levels of
all the nodes in the network to obtain the output. Then,
starting at the output nodes, a backward pass is used to
calculate ∂ET ∂ω , where ω denotes the adjustable
parameters in the network. The general parameter
update rule is given by
 ∂ET 
ω (t + 1) = ω (t ) + ∆ω (t ) = ω (t ) + η  −  , (11)
 ∂ω (t ) 
where η denotes the learning rate.
The normal vector calculated from the network is
denoted as n k = (v1k , v 2k , v3k ) T for the kth point on
the surface. The normal vectors n k is updated

3
406

Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.
iteratively using the gradient method as:
v jk (t + 1) = v jk (t ) + 2η s j (t )[Dk (t ) − Rk (t ) ], j = 1, 2, 3,
(12)
where s j (t ) denotes the jth element of illuminant
direction s; The updated v jk should be normalized as
follows:
v jk (t + 1)
v jk (t + 1) = , j = 1, 2, 3. (13)
n k (t + 1)
Since the structure of the proposed neural
networks is like a mirror in the center layer, the update
rule for the weights between layer 2 and layer 3 of the
networks denoted as W (see Fig. 1) can be calculated
by the least square method. Hence, W at time t + 1
can be calculated by
( T
W (t + 1) = V (t + 1) V (t + 1) )
−1
V (t + 1) ,
where V (t + 1) denotes the weights betweens the
output and central layers of the mirror-networks.
Finally, the values n k (t + 1) are denoted as
albedo αk , where k = 1,2,..., m.
5. 3D Surface Reconstruction from the Surface
Normal by Enforcing Integrability
The enforcing integrability approach was proposed
in the earliest stage by R. T. Frankot and R. Chellappa
[24]. Suppose that we represent the surface z ( x, y ) by
a finite set of integrable basis functions φ (x, y, ω ) so
that
z (x, y ) = ∑ c(ω )φ (x, y , ω ) , (15)
ω∈Ω
where ω = (u, v ) is a two-dimensional index, Ω is a
finite set of indexes, and { φ (x, y, ω ) } is a finite set of
integrable basis functions which are not necessarily
mutually orthogonal. We chose the discrete cosine
basis so that { c (ω) } is exactly the full set of discrete
cosine transform (DCT) coefficients of z ( x, y ) . Note
that the partial derivatives of z ( x, y ) can also be
expressed in terms of this expansion, giving
z x (x, y ) = ∑ c(ω ) φx (x, y, ω ) (16)
ω∈Ω
z y ( x, y ) = ∑ c(ω)φ y (x, y, ω ) ,
ω∈Ω
where φ x ( x, y, ω ) = ∂φ (⋅) ∂x and φ y (x, y, ω ) = ∂φ (⋅) ∂y .
Suppose we now have the possibly nonintegrable
estimate n( x, y ) from which we can easily deduce
from Eq. (2) the possibly nonintegrable partial
Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.
derivatives zˆ x ( x, y ) and zˆ y ( x, y ) . These partial
derivatives can also be expressed as a series, giving
zˆx ( x, y) = ∑cˆ1(ω) φx (x, y, ω) (18)
ω∈Ω
zˆy (x, y) = ∑cˆ2 (ω)φy (x, y, ω) . (19)
ω∈Ω
A method has been proposed for finding the
expansion coefficients c(ω) given a possibly
nonintegrable estimate of surface slopes zˆ x ( x, y ) and
zˆ y ( x, y ) :
p x (ω )cˆ1 (ω ) + p y (ω )cˆ2 (ω )
c(ω ) = , for ω = (u, v ) ∈ Ω , (20)
p x (ω ) + p y (ω )
px (ω) = ∫∫ φx ( x, y, ω) dxdy
2
where and
2
p y (ω ) = ∫∫ φ y (x, y, ω ) dxdy . Finally, we can reconstruct
T
(14) the object’s surface by performing the inverse 2-D
DCT on the coefficients c(ω) .
6. Experimental Results and Discussions
Quantitative experimental results have been
obtained by reconstructing a synthetic sphere object.
The true depth map of the synthetic sphere object is
generated mathematically as
 r 2 − x 2 − y 2 , if x 2 + y 2 ≤ r 2
z (x, y ) =  ,(21)
 0, otherwise
where r=48, 0 < x , y ≤ 100 , and the center is located at
(x, y)=(51, 51). The sphere object is showed in Fig. 2.
This synthetic image was generated using the depth
function in Eq. (21) and the surface gradients were
computed using the discrete approximation. Figure 3
shows the synthetic images generated according to the
Lambertian model with varying albedo and different
directions. The different albedos are, 0.6 for
right-bottom of the sphere, 0.8 for left-top of the
sphere, and 1 for the rest part. The locations of light
sources in Figs. 3(a)-(i) are S1=(30, 140), S2=(30, 90),
S3=(30, 40), S4=(30, 180), S5=(0, 0), S6=(30, 0),
S7=(30, -140), S8=(30, -90), and S9=(30, -40), where
the first component is the degree of tilt angle and the
second component is the degree of pan angle. The
center of image is set as the origin of the coordination.
(17) The x-y plane is parallel to the image plane. The z-axis
is perpendicular to the image plane. The experimental
results are shown in Table 1 and the proposed method
is compared with two photometric stereo algorithms,
Hayakawa’s method [19] and Georghiades’s method
[10]. In Table 1, we take 5 groups of images with
different illuminant angles from the left, the right, and
4
407
the front for 3D reconstruction. Both estimated surface In the second experiment, we test the algorithm
and synthetic one are normalized within the interval [0, on a number of real images from the Yale Face
1]. According to Table 1, it is found that the proposed Database B [25] showing the variability due to
method can achieve the lowest mean errors compared illumination and there is varying albedo in each point
with the other methods in all illumination conditions. of surface of human faces.
First, we take the images of the same person that
was photographed under three different light sources
from these testing images arbitrarily shown in Fig.
4(a)-(c). We feed the normalized images into our
algorithm. After updating the parameters by several
iterations, we can get the normal vector of the surfaces
of human faces corresponding to each pixel in an
image in the weights of networks. The results are
shown in the second row in Fig. 4, which are the first
component, the second component, and the third
component of the surface normal vector in order.
Figure 2. Synthetic sphere surface object.
Figure 5 presents the results of 3D human face
reconstruction. Figure 5(a) shows the surface albedo of
human face in Fig. 4. Figure 5(b) shows the result with
our proposed algorithm. By using the Georghiades’s
approach [10] and the Hayakawa’s approach [19], the
reconstructed results are demonstrated in Fig. 5(c) and
(a) S1=(30, 140) (b) S2=(30, 90) (c) S3=(30, 40)
5(d), respectively. The results clearly indicate that the
performance of our proposed nonlinear reflectance
model is better than that of the Georghiades’s approach
and the Hayakawa’s approach. Comparing to the
results obtained by the Georghiades’s approach, the
reconstructed surfaces with the consideration of
(d) S4=(30, 180) (e) S5=(0, 0) (f) S6=(30, 0) specular components in our algorithm, are obviously
better in high-gradient parts such as the nose. Besides,
the Hayakawa’s approach did need added constraints,
it could reconstruct the 3D model of human face
similar as our approach, but when the constraints are
unavailable, then it could not reconstruct the 3D model
(g) S7=(30, -140) (h) S8=(30, -90) (i) S9=(30, -40)
Figure 3. The 2D sphere images generated with varying
of human face.
albedo and different lighting directions (the
degree of tilt angle, the degree of pan angle).

Table 1. The absolute mean errors between estimated

depths and desired depth of synthetic object’s 3D
surfaces. (Both light and viewing directions are
unknown in the experiment.) (a) (b) (c)
Mean Georghiades’ Our
Hayakawa’s
absolute Lights s method proposed
method ([19])
depth error ([10]) method
S1, S2, S3 0.049 0.077 0.020
Sphere with S7, S8, S9 0.050 0.079 0.027
Variant S1, S5, S3 0.033 0.073 0.021
albedo S1, S8, S6 0.034 0.079 0.018 (d) (e) (f)
S1, S5, S7 0.032 0.076 0.018 Figure 4. (a)-(c)Three training images with differ light source
positions from Yale Face Database B in frontal. (d)-(f)
Surface normal corresponding to the three source
images.

5
408

Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.
 Acknowledgment
This work was supported by the NSC, under
94-2622-E-164-001-CC3.
(a) (b)
(c) (d)
Figure 5. (a)The surface albedo of human face in Fig. 4. The
results of 3D model reconstruction by (b) our
proposed algorithm, (c) Georghiades’s approach in
[10], and (d) Hayakawa’s approach in [19].
7. Conclusions
In this paper, we used the images under three
different light source locations to solve this problem
which have to know the locations of light sources first
for solving the photometric stereo based on Lambertian
model. In our approach, it can obtain a very good
result even if the locations of light sources are not
given. Addition, using the structure of the proposed
symmetric neural network for solving photometric
stereo problems does not need any other desired output
value and the smoothing conditions. It is easier to
converge and make the system stable.
The performance comparisons of our proposed
neural-network-based photometric stereo approach to
the Georghiades’s approach in [10] and the
Hayakawa’s approach in [19] were made. We test the
algorithm on synthetically generated images for the
reconstruction of surface of objects. The results clearly
indicate that the performances of our proposed
approach are better than that of the Georghiades’s
approach in [10] and the Hayakawa’s approach in [19].
All the experimental results showed that the
performance of the proposed approach is better than
those of the two proposed existing photometric stereo
approaches.
6
409
Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.
References
[1] R. J. Woodham, “Photometric Method for Determining
Surface Orientation from Multiple Images,” Jour. of
Optical Engineering, Vol. 19, No. 1, 1980.
[2] K. M. Lee and C. J. Kuo, “Shape Reconstruction from
Photometric Stereo,” Jour. of Optical Society of
America: A, Vol. 10, No. 5, 1993.
[3] C. Cho and H. Minanitani, “A New Photometric
Method Using 3 Point Light Sources,” IEICE Trans.
Inf. and Syst. V.E76-D, No. 8, pp. 898-904, August
1993.
[4] Y. Iwahori, R. Woodham, and A. Bagheri, “Principal
Components Analysis and Neural Network
Implementation of Photometric Stereo,” Proc. of the
Workshop on Physics-Based Modeling in Computer
Vision, pp. 117-125, June 1995.
[5] G. Kay and T. Caelli, “Estimating the Parameters of an
Illumination Model Using Photometric Stereo,”
Graphical Models and Image Processing, Vol. 57, No.
5, pp. 365-388, 1995.
[6] G. McGunnigle, “The Classification of Textured
Surfaces under Varying Illuminant Direction,” Ph.D.
Thesis, Department of Computing and Electrical
Engineering, Heriot-Watt University, Edinburgh, 1998.
[7] S. K. Nayar, K. Ikeuchi, and T. Kanade, “Determining
Shape and Reflectance of Hybrid Surfaces by
Photometric Sampling,” IEEE Trans. on Robotics and
Automation, Vol. 6, No. 4, pp. 418-431, 1990.
[8] E. Angelopoulou and J. P. Williams, “Photometric
Surface Analysis in a Triluminal Environment,” IEEE
International Conf. on Computer Vision, 1999.
[9] P. N. Belhumeur, D. J. Kriegman, and A. L. Yuille,
“The Bas-Belief Ambiguity,” IEEE Conf. on CVPR,
pp. 1060-1066, 1997.
[10] S. Georghiades, P. N. Belhumeur, and D. J. Kriegman,
“From Few to Many: Illumination Cone Models for
Face Recognition under Variable Lighting and Pose,”
IEEE Trans. on PAMI, Vol. 23, No. 6, pp. 643-660,
June 2001.
[11] S. Georghiades, “Incorporating the Torrance and
Sparrow Model of Reflectance in Uncalibrated
Photometric Stereo,” IEEE International Conf. on
Computer Vision, 2003.
[12] K. Ikeuchi, “Determining Surface Orientations of
Specular Surfaces by Using the Photometric Stereo
Method,” IEEE Trans. on PAMI, Vol. 3, pp. 661-669,
1981.
[13] E. N. Coleman and R. Jain, “Obtaining 3Dimensional
Shape of Textured and Specular Surfaces Using
Four-Source Photometry,” Proc. of Computer Vision,
Graphics, and Image, Vol. 18, pp. 309-328, 1982.
[14] H. D. Tagare and R. J. P. deFigueiredo, “Simulataeous
Estimation of Shape and Reflectance Maps from
Photometric Stereo,” IEEE International Conf. on
Computer Vision, pp. 340-343, 1990.
[15] F. Solomon and K. Ikeuchi, “Extracting the Shape and
Roughness of Specular Lobe Objects Using Four Light
Photometric Stereo,” IEEE Conf. on CVPR, pp.
466-471, 1992.
[16] H. Rushmeier, G. Taubin, and A. Gueziec, “Applying
Shape from Lighting Variation to Bump Map
Capture,” In Eurographics Rendering Techniques’97,
pp. 35-44, St. Etienne, France, June 1997.
[17] O. Drbohlav and A. Leonardis, “Detecting Shadows
and Specularities by Moving Light,” Proc. of
Computer Vision Winter Workshop, Ljubljana,
Slovenia, pp. 60-74, 1998.
[18] O. Drbohlav and R. Sara, “Specularities Reduce
Ambiguity of Uncalibrated Photometric Stereo,” IEEE
Conf. on Computer Vision, Copenhagen, Denmark,
2002.
[19] K. Hayakawa, “Photometric Stereo under a Light
Source with Arbitrary Motion,” Jour. of the Optical
Society of America: A, Vol. 11, No. 11, 1994.
[20] G. Q. Wei and G. Hirzinger, “Learning shape from
shading by a multilayer network,” IEEE Trans. on
Neural Networks, Vol. 17, pp. 985-995, 1996.
[21] S.Y. Cho and T. W. S. Chow, “Shape recovery from
shading by a new neural-based reflectance model,”
IEEE Trans. on Neural Networks, Vol. 10,
pp.1536-1541, 1999.
[22] S. Y. Cho and T. W. S. Chow, “Neural computation
approach for developing a 3D shape reconstruction
model,” IEEE Trans. on Neural Networks, Vol. 12,
No. 5, September 2001.
[23] R. D. Morris, P. Cheeseman, V. N. Smelyanskiy, and
D. A. Maluf, “A Bayesian Approach to High
Resolution 3D Surface Reconstruction from Multiple
Imagesm,” Proc. of the IEEE Signal Processing
Workshop on, 14-16 June 1999, pp. 140 – 143.
[24] R. T. Frankot and R. Chellappa, “A Method for
Enforcing Integrability in Shape From Shading
Algorithms,” IEEE Trans. on PAMI, Vol. 10, No. 4,
pp. 439-451, July 1988.
[25] Online available =
http://cvc.yale.edu/projects/yalefacesB/yalefacesB.htm
l.

7
410

Authorized licensed use limited to: Corporacion Universitaria de la Costa. Downloaded on April 20,2023 at 16:05:53 UTC from IEEE Xplore. Restrictions apply.