A Multiresolution Spline With Application to
`Image Mosaics
`
`PETER J. BURT and EDWARD H. ADELSON
`
`RCA David Sarnoff Research Center
`
`We define a multiresolution spline technique for combining two or more images into a larger image
`mosaic. In this procedure,
`the images to be splined are first decomposed into a set of band-pass
`filtered component images. Next, the component images in each spatial frequency hand are assembled
`into a corresponding bandpass mosaic. In this step, component images are joined using a weighted
`average within a transition zone which is proportional in size to the wave lengths represented in the
`band. Finally, these band-pass mosaic images are summed to obtain the desired image mosaic. In this
`way, the spline is matched to the scale of features within the images themselves. When coarse features
`occur near borders, these are blended gradually over a relatively large distance without blurring or
`otherwise degrading finer image details in the neighborhood of th e border.
`
`Categories and Subject Descriptors: I.3.3 [Computer Graphics]: Picture/Image Generation;
`[Image Processing]: Enhancement
`
`I.4.3
`
`General Terms: Algorithms
`
`Additional Key Words and Phrases: Image mosaics, photomosaics, splines, pyramid algorithms,
`multiresolution analysis, frequency analysis, fast algorithms
`
`1. INTRODUCTION
`
`The need to combine two or more images into a larger mosaic has arisen in a
`number of contexts. Panoramic views of Jupiter and Saturn have been assembled
`for multiple
`images returned to Earth from the two Voyager spacecraft. In a
`similar way, Landsat photographs are routinely assembled into panoramic views
`of Earth. Detailed images of galaxies and nebulae have been assembled from mul-
`
`The work reported in this paper was supported by NSF grant ECS-8206321. A shorter description of
`this work was published in the Proceedings of SPIE, vol. 432, Applications of Digital Image Processing
`VI, The International Society for Optical Engineering, Bellingham, Washington.
`Authors' address: RCA David Sarnoff Research Center, Princeton, NJ 08540.
`Permission to copy without fee all or part of this material is granted provided that the copies are not
`made or distributed for direct commercial advantage, the ACM copyright notice and the title of the
`publication and its date appear, and notice is given that copying is by permission of the Association
`for Computing Machinery. To copy otherwise, or to republish, requires a
`fee and/or specific
`permission.
`© 1983 ACM 0730-0301/83/1000-0217 $00.75
`
`ACM Transactions on Graphics, Vol. 2. No. 4, October 1983, Pages 217-236.
`
`VALEO EXHIBIT 1025
`Valeo v. Magna
`IPR2015-____
`
`VALEO EX. 1025_001
`
`

`
`218 • P. J. Burt and E. H. Adelson
`
`Fig. 1. A pair of images may be represented as a pair of surfaces above the (x, y) plane. The problem
`of image splining is to join these surfaces with a smooth seam, with as little distortion of each surface
`as possible.
`
`is
`technique
`the mosaic
`these cases,
`tiple telescope photographs. In each of
`used to construct an image with a far larger field of view or level of detail than
`could be obtained with a single photograph. In advertising or computer graphics,
`the technique can be used to create synthetic images from possibly unrelated
`components.
`A technical problem common to all applications of photomosaics is joining two
`images so that the edge between them is not visible. Even slight differences in
`image gray level across an extended boundary can make
`that boundary quite
`noticeable. Unfortunately, such gray level differences are frequently unavoidable;
`they may be due to such factors as differences in camera position or in image
`processing prior to assembly. Thus, a technique is required which will modify
`image gray levels in the vicinity of a boundary to obtain a smooth
`transition
`between images. The two images to be joined may be considered as two surfaces,
`where the image intensity I(x, y) corresponds to
`the elevation above the x,y
`plane. The problem, as illustrated in Figure 1, may be stated as follows: How can
`the two surfaces be gently distorted so that they can be joined together with a
`smooth seam? We will use the term image spline to refer to digital techniques for
`making these adjustments. A good image spline will make the seam perfectly
`smooth, yet will preserve as much of the original image information as possible.
`It is probably safe to say that no fully satisfactory splining technique has yet
`been found. Most image mosaics are still produced without any attempt at remov-
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October l983.
`
`VALEO EX. 1025_002
`
`

`
`A Multiresolution Spline with Application to Image Mosaics • 219
`
`Fig. 2. The weighted average method may be used to avoid seams when mosaics are constructed from
`overlapped images. Each image is multiplied by a weighting function which decreases monotonically across
`its border; the resulting images are then summed to form the mosaic. Example weighting functions are shown
`here in one dimension. The width of the transition zone T is a critical parameter for this method.
`
`level difference
`ing visible boundaries (e.g., [4]). The magnitude of the gray
`across a mosaic boundary can be reduced to some extent by a judicious choice of
`boundary location when splining overlapped images. The match may be improved
`by adding a linear ramp to pixel values on either side of the boundary to obtain
`equal values at the boundary itself [6, 7]. A still smoother transition can be
`obtained using a
`technique recently proposed by Peleg [9]: The "smoothest
`possible" correction function is constructed which can be added to each image of
`a mosaic to eliminate edge differences. However,
`this
`technique may not be
`practical for large images, since the correction functions must be computed using
`an iterative relaxation algorithm.
`We are concerned with a weighted average splining technique. To begin, it is
`assumed that the images to be joined overlap so that it is possible to compute
`the gray level value of points within a transition zone as a weighted average of
`the corresponding points in each image. Suppose that one image, Fl(i), is on the
`left and the other, Fr(i), is on the right, and that the images are to be splined at
`a point ˆi (expressed in one dimension to simplify notation). Let Hl(i) be a
`weighting function which decreases monotonically from
`left
`to right and
`let
`Hr(i) = 1 – Hl(i) (see Figure 2). Then, the splined image F is given by
`
`F(i) = Hl(i— ˆi ) Fl(i) + Hr(i— ˆi ) Fr(i).
`
`It is clear that with an appropriate choice of H, the weighted average technique
`will result in a transition which is smooth. However, this alone does not ensure
`that the location of the boundary will be invisible. Let T be the width of a
`transition zone over which Hl changes from 1 to 0. If T is small compared to
`image features, then the boundary may still appear as a step in image gray level,
`albeit a somewhat blurred step. If, on the other hand, T is large compared to
`image features, features from both images may appear superimposed within
`the
`transition zone, as in a photographic double exposure.
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_003
`
`

`
`220 • P. J. Burt and E. H. Adelson
`
`These extremes are illustrated in Figure 3 with several attempts to spline two
`synthetic images of stars. The original images, Figures 3a and 3b (257 x 257
`pixels) are identical except for a slight shift in vertical position and a slight shift
`in mean gray level. The first of these differences can arise from optical distortions
`or misalignments of actual photographs, while the second can be due to differ-
`ences in atmospheric conditions or in photographic development.
`In this example, photomosaics are obtained by joining the left half of Figure
`3a with the right half of Figure 3b. If this is done without any attempt to smooth
`the image transition (T = 0), the boundary will appear as a sharp edge (Figure
`3c). If instead the images are combined by the method of weighted average within
`a relatively narrow transition zone (T = 8), the edge appears blurred but
`remains
`visible (Figure 3d). When the images are splined with a broad transition (T =
`64), the edge is no longer visible but stars have a "double exposed" look within
`the transition zone (Figure 3e).
`Clearly, the size of the transition zone, relative to the size of image features,
`plays a critical role in image splining. To eliminate a visible edge the transition
`width should be at least comparable in size to the largest prominent features in
`the image. On the other hand, to avoid a double exposure effect, the zone should
`not be much larger than the smallest prominent image features. There is no
`choice of T which satisfies both requirements in the star images of Figure 3
`because these contain both a diffuse background and small bright stars.
`These constraints can be stated more precisely in terms of the image spatial
`frequency content. In particular, a suitable T can only be selected if the images
`to be splined occupy a relatively narrow spatial frequency band. As a rough
`requirement, we may stipulate that T should be comparable in size to the wave-
`length of the lowest prominent frequency in the image. If T is smaller than this
`the spline will introduce a noticeable edge. On the other hand, to avoid a double
`exposure effect, T should not be much larger than two wave lengths of the highest
`prominent
`frequency component in the images. This ensures that
`there will not
`be room for multiple features within the transition zone. While it is likely that
`these
`limits
`can be exceeded somewhat without noticeable degradation,
`the
`general conclusion—that the band width of
`images
`to be splined should be
`roughly one octave—is an important one.
`How can images which occupy more than an octave be splined? The approach
`proposed here is that such images should first be decomposed into a set of band-
`pass component images. A separate spline with an appropriately selected T can
`then be performed in each band. Finally, the splined band-pass components are
`recombined into the desired mosaic image. We call this approach the multi-reso-
`lution spline. It was used to obtain the image shown in Figure 3f.
`In decomposing the image into frequency bands, it is important that the range
`of frequencies in the original be covered uniformly, although the bands themselves
`may overlap. As a practical matter, a set of low-pass filters are applied to generate
`a sequence of images in which the band limit is reduced from image to image in
`one-octave steps. Band-pass images can then be obtained simply by subtracting
`each low-pass image from the previous image in
`the sequence. This not only
`ensures complete coverage of spatial
`frequencies but means that the final mosaic
`can be obtained simply by summing the band-pass component images.
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_004
`
`

`
`A Multiresolution Spline with Application to Image Mosaics • 221
`
`Fig. 3. Common artifacts of the weighted average techniques are demonstrated in these attempts to
`spline two synthetic images of stars (Figure 3a and 3b). These differ only in mean gray level and a
`slight vertical shift. A seam is clearly visible when the left half of figure 3a is joined with the right
`half of Figure 3b without any adjustment in gray level, as shown in Figure 3c. The seam is still visible
`when the weighted average technique is used with a narrow transition zone (Figure 3d). However, if
`the transition zone is wide, features within the zone appear double (Figure 3e). The first of these
`artifacts is due to a gray level mismatch at low spatial frequencies, while the second is due to a
`position mismatch at high frequencies. Both are avoided in the multiresolution method (Figure 3f).
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_005
`
`

`
`222 • P. J. Burt and E. H. Adelson
`
`Fig. 4. A one-dimensional graphical representation of the iterative REDUCE operation used
`pyramid construction.
`
`in
`
`In Section 2 we present a highly efficient ''pyramid'' algorithm for performing
`the required filtering operations and in Section 3 we show that
`the pyramid
`structure is ideally suited for performing the splining steps as well.
`
`2. BASIC PYRAMID OPERATIONS
`
`A sequence of low-pass filtered images G0, G1, . . ., GN can be obtained by
`repeatedly convolving a small weighting function with an image [1, 3]. With this
`technique, image sample density is also decreased with each iteration so that the
`bandwidth is reduced in uniform one-octave steps. Sample reduction also means
`that the cost of computation is held to a minimum.
`Figure 4 is a graphical representation of the iterative filtering procedure in one
`dimension. Each row of dots represents the samples, or pixels, of one of the
`filtered images. The lowest row, G0, is the original image. The value of each node
`in the next row, G1, is computed as a weighted average of a 5 x 5 subarray of G0
`nodes, as shown Nodes of array G2 are then computed from G1 using the same
`pattern of weights. The process is iterated to obtain G2 from G1, G3 from G2 and
`so on. The sample distance is doubled with each iteration so that successive
`arrays are half as large in each dimension as their predecessors. If we imagine
`these arrays stacked one above the other, the result is the tapering data structure
`known as a pyramid [10]. If the original image measures 2N + 1 by 2N + 1, then
`the pyramid will have N + 1 levels.1
`level of the
`Both sample density and resolution are decreased from level to
`pyramid. For this reason, we shall call the local averaging process which generates
`each pyramid level from its predecessor a REDUCE operation. Again, let G0 be
`the original image. Then for 0 < l < N:
`
`by which we mean:
`
`Gl = REDUCE [Gl-1],
`
`(
`
`2
`
`+
`+
`
`
`
`i m j n,2
`
`).
`
`−
`1
`
`5
`
`,
`m n
`
`=∑
`
`w m n G( , )
`
`
`
`
`l
`1
`
`G i j( , )
`
`
`l
`
`=
`
`∑
`
`________
`More generally, a pyramid of N + 1 levels may be constructed from any image which measures
`MR2N + 1 rows by MC2N + 1 columns, where MR and MC are integers.
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_006
`
`

`
`A Multiresolution Spline with Application to Image Mosaics • 223
`
`The pattern of weights w(m, n) used to generate each pyramid level from its
`predecessor is called the generating kernel. These weights are chosen subject to
`four constraints: First,
`for computational convenience, the generating kernel
`is
`ˆw (m)
`ˆw (n). Second, the one dimensional function
`ˆw is
`separable, w(m, n) =
`ˆw (0) = a,
`ˆw ( - 1) = ˆw (1) = b, and ˆw (- 2) = ˆw (2) = c, as shown in
`symmetric,
`Figure 4. Third, ˆw is normalized, a + 2b + 2c = 1. The final constraint stipulates
`that each level l node must contribute the same total weight to level l + 1 nodes:
`thus, a + 2c = 2b. Now, combining constraints, we find that a may be considered
`a free variable, while b = 1/4 and c = 1/4 — a/2.
`
`2.1 Equivalent Weighting Functions
`
`It is clear that every level l node in the pyramid represents a weighted average of
`a 5 × 5 subarray of level l – 1 nodes. Each of these in turn represents an average
`of a subarray of level l – 2. In this way, we can trace the weights for a given
`pyramid node back to the original image G0 to discover the "equivalent weighting
`function" Wl which, if convolved directly with the original image, would have
`given the same node values at level l. It is convenient to discuss pyramid-based
`computations
`in
`terms of
`these equivalent weighting functions, although
`the
`iterative REDUCE process is considerably more efficient and
`is used
`in all
`computations.
`The equivalent weighting functions have several properties which will be
`important
`in filtering and splining operations. The scale of
`these
`functions
`doubles from level to level of the pyramid while their shape does not change [1].
`Function shape does depend on the value of parameter a in the generating kernel.
`For example, if a = 0.5, the functions are all triangular in shape, while if a = 0.4,
`the functions resemble the Gaussian probability density function. Convolution
`with a Gaussian has the effect of low-pass filtering the image. Pyramid construc-
`tion is equivalent to convolving the image with a set of Gaussian-like functions
`to produce a corresponding set of filtered images. Because of the importance of
`the multiple filter interpretation, we shall refer to this sequence of images G0,
`G1, . . ., GN as the Gaussian pyramid.
`Suppose samples in G0 are separated by a unit distance. Then, samples at level
`l are separated by the distance 2 l. It can be shown that the width of the equivalent
`weighting function W l is 2l+2 – 4, covering 2l+2 – 3 image samples, or just less
`than four times the sample distance (see Figure 5). Thus equivalent weighting
`functions centered on level l sample points will overlap in such a way that each
`image pixel contributes to the value of at most 16 level-l samples (4 in one
`dimension). If the contributions of any image pixel are summed, the result will
`be unity. For each i, j, and l,
`
`2
`
`∑
`
`
`
`m n,
`
`−
`l
`
`W i m j( ,2
`l
`=−
`2
`
`
`
`
`
`−
`
`l
`
`n
`
`2
`
`)
`
`=
`
`
`
`.1
`
`This result follows from the equal contribution property of the generating kernel.
`The Gaussian shape and summation properties of
`the functions W l mean
`they can be used to construct the weighting functions H needed for image splining
`
`ACM Transactions on Graphics, Vol. 2, No 4, October 1983.
`
`VALEO EX. 1025_007
`
`

`
`224 • P. J. Burt and E. H. Adelson
`
`Fig. 5. Equivalent weighting functions, Wl, are shown centered at level l sample points on the left
`in the figure, while zero weight is given to points on the right. When these weights are summed, a
`uniform value of unity is obtained on the left, a value of zero on the right, and a monotonic transition
`in the center. The weighting functions H used in image splining can be constructed in this way (see
`Figure 2).
`
`(Figure 2). Suppose Wl is associated with each node in the left half of Gl while
`zero weight is associated with nodes in the right half (Figure 5). Then, the sum
`of these functions will be a function which decreases monotonically from 1 to 0,
`with a transition zone width T equal to 3 times the level l sample interval. This
`property will be used
`in
`the pyramid-based multiresolution spline, although
`functions H and W will never be explicitly computed.
`
`2.2 The Laplacian Pyramid
`
`The Gaussian pyramid is a set of low-pass filtered images. In order to obtain the
`band-pass images required for the multiresolution spline we subtract each level
`of the pyramid from the next lowest level. Because these arrays differ in sample
`density, it is necessary to interpolate new samples between those of a given array
`before it is subtracted from the next lowest array. Interpolation can be achieved
`by reversing the REDUCE process. We shall call this an EXPAND operation.
`Let Gl,k be the image obtained by expanding Gl k times. Then
`
`and for k > 0,
`
`Gl,0 = Gl,
`
`Gl,k = EXPAND[Gl,k-1].
`
`.
`
`
`
`n
`
`2
`
`i m j2
`
`+
`
`,
`2
`
`+
`2
`
`
`
`G
`
`
`l k,
`
`−
`
`1
`
`By EXPAND, we mean
`
`G i j( , )
`
`
`
`l k,
`
`=
`
`2
`
`∑∑4
`
`
`
`n n,
`
`=−
`2
`
`Here, only terms for which (2i + m)/2 and (2j + n)/2 are integers contribute
`to the sum. Note that Gl,1 is the same size as Gl-1, and that G l,l is the same size as
`the original image.
`
`ACM Transactions on Graphics, Vol. 2, No. 4. October 1983.
`
`VALEO EX. 1025_008
`
`

`
`A Multiresolution Spline with Application to Image Mosaics • 225
`
`We now define a sequence of band-pass images L0, L1, . . ., LN. For 0 < l < N,
`
`Ll = Gl - EXPAND[Gl+1] = Gl - Gl+l,l.
`
`Because there is no higher level array to subtract from GN, we define LN = GN.
`Just as the value of each node in
`the Gaussian pyramid could have been
`obtained directly by convolving the weighting function W l with the image, each
`node of L l can be obtained directly by convolving W l - Wl+1 with the image. This
`difference
`of Gaussian-like
`functions
`resembles
`the Laplacian
`operators
`commonly used
`in
`the
`image processing [5], so we refer
`to
`the
`sequence
`L0, L1, . . ., LN as the Laplacian pyramid.2
`
`2.3 Summation Property
`
`The steps used to construct the Laplacian pyramid may be reversed to recover
`the original image G0 exactly. The top pyramid level, L N, is first expanded and
`added to LN-1 to recover GN-1; this array is then expanded and added to LN-2 to
`recover GN-2, and so on. Alternatively, we may write
`
`N
`
`
`
`Ll l, .
`
`=∑
`
`l
`
`0
`
`G
`0
`
`=
`
`The expand and sum procedure will be used to construct a mosaic image from
`its set of splined band-pass components.
`
`2.4 Boundary Conditions
`
`In both the REDUCE and EXPAND operations, special attention must be given
`to edge nodes. For example, when a REDUCE is performed, the generating kernel
`for an edge node at level Gl=1 extends beyond the edge of level Gl by two nodes.
`Therefore, before the REDUCE (or EXPAND) is performed, Gl is augmented by
`two rows of nodes on each side. Values are assigned to these nodes by reflection
`and inversion across the edge node. Thus, if Gl(0, j) is a node on the left edge of
`Gl, we set
`
`and
`
`Gl( -1, j) = 2Gl(0, j) - Gl(1, j),
`
`Gl( -2, j) = 2Gl(0, j) - Gl(2, j).
`
`This treatment of boundaries has the effect of extrapolating the images in such
`a way that the first derivative is constant at the edge node (the second derivative
`is zero).
`
`3. THE MULTIRESOLUTION SPLINE
`
`3.1 Splining Overlapped Images
`
`terms of
`The multiresolution spline algorithm may be defined rather simply in
`the basic pyramid operations introduced in
`the last section. Here variations on
`the method will be described for splining overlapped and nonoverlapped square im-
`_________
`2In fact, the equivalent weighting functions for the Laplacian pyramid are slightly different from
`l + 1.
`Wl - Wl-1, because of the EXPAND operation used at level
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_009
`
`

`
`226 • P. J. Burt and E. H. Adelson
`
`images of arbitrary shape. Modifications for other tasks
`ages and for splining
`will then be apparent. To begin, suppose we wish to spline the left half of image
`A with the right-half of image B. Assume that these images are both square,
`measuring 2N + 1 pixels on a side, and that they overlap completely. The spline
`is achieved in three steps:
`
`Step 1. Laplacian pyramids LA and LB are constructed for images A and B
`respectively.
`Step 2. A third Laplacian pyramid LS is constructed by copying nodes from
`the left half of LA to the corresponding nodes of LS, and nodes in the
`right half of LB to the right half of LS. Nodes along the center line of
`LS are set equal to the average of corresponding LA and LB nodes.
`The center line for level l of a Laplacian pyramid is at i = 2N-1. Thus, for
`all
`i,
`j,
`l,
`
`− − −
`
`1
`
`1
`
`1
`
`N
`
`N
`
`N
`
`2 2 2
`< = >
`
`1
`
`i i
`
`if
`
`if
`
`if
`
`+
`
`LB i j( , )) /
`
`
`l
`
`2
`
`  
`
`LS i j( , )
`
`
`l
`
`=
`
`(
`
`LA i j( , )
`
`
`l
`
`LA i j( , )
`
`
`l
`LB i j( , )
`
`
`l
`
`Step 3. The splined image S is obtained by expanding and summing the levels
`of LS.
`
`The result of applying this procedure to the star example is shown in Figure
`3f. Note that the transition between image halves is now smooth, without the
`blurred step edge of Figure 3d or the doubling of Figure 3e.
`A second example is given in Figure 6. Here we wish to spline two Landsat
`images of San Francisco, Figures 6a and 6b. These images are identical except
`for diffuse background noise which has been added to simulate the effects of
`possible differences in atmospheric conditions or image processing. Again, we
`wish to construct a mosaic in which the left half of one image is joined to the
`right half of the other. If this is done without a spline, the boundary is easily
`visible (Figure 6c). If the multiresolution spline is used, however, the edge is
`completely removed (Figure 6d).
`images, an
`A third example shows the result of splining two quite different
`apple and an orange (Figures 7a and 7b). The mosaic obtained without a spline
`is given in Figure 7c, while that obtained with the spline is shown in Figure 7d.
`In this case the transition between component images has been made slightly
`more gradual: in addition to averaging Laplacian nodes along the center
`line,
`nodes on either side of the center nodes have been averaged with a 3/4 to 1/4
`ratio of weights. The splining process has been repeated separately for the red,
`green, and blue image color components. Again, a smooth transition is obtained
`despite the rather large step in gray level between the apple and orange halves.
`In this pyramid-based splining procedure,
`the equivalent weighting functions
`W l play a dual role. Within the domain of each image, they act as interpolation
`functions between level l samples. Along the boundary between the image halves,
`they act as the splining functions H. If the images to be splined are identical,
`then the mosaic obtained through the pyramid-based splining will be the same im-
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_010
`
`

`
`A Multiresolution Spline with Application to Image Mosaics • 227
`
`Fig. 6. The spline applied to Landsat images of San Francisco. When the left half of Figure 6a is joined to the
`right half of Figure 6b without a spline, the boundary is clearly visible (Figure 6c). No boundary is visible
`when the multiresolution spline is used (Figure 6d).
`
`introduce
`age again. In this sense, the splining procedure by itself does not
`image distortion. As shown in Figure 5, W l extends twice the sample distance 2l
`on each side of the level l sample point. This is an appropriate transition distance
`for splining the frequencies represented in the lth pyramid level.
`
`3.2 Splining Regions of Arbitrary Shape
`
`The steps outlined above can be generalized for constructing a mosaic from image
`regions of arbitrary shape. Again, we assume that the regions to be splined are
`contained in images A and B and that these completely overlap. As before, nodes
`of the Laplacian pyramids LA and LB for the component images will be combined
`to form the Laplacian pyramid LS of the image mosaic S . We introduce an
`additional pyramid structure in order to determine which nodes of LS should be
`taken from LA, which from LB, and which should be an average of the two. Let
`R be a binary image of the same size as A and B, in which all pixels inside the
`region of A to be splined with B are 1 and all those outside the region are 0. The
`steps of the multiresolution spline are modified as follows (on page 230):
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_011
`
`

`
`228 • P. J. Burt and E. H. Adelson
`
`Fig. 7. The spline can be used to combine very different images. Here the left half of an apple
`(Figure 7a), is combined with the right half of an orange (Figure 7b). Figure 7c, obtained without a
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983
`.
`
`VALEO EX. 1025_012
`
`

`
`A Multiresolution Spline with Application to Image Mosaics • 229
`
`spline, shows that the orange and apple differ considerably in gray level and color. Still, a smooth
`transition is obtained with the multiresolution spline (Figure 7d).
`
`ACM Transactions on Graphics, Vol. 2, No. 4, October 1983.
`
`VALEO EX. 1025_013
`
`

`
`230 • P. J Burt and E. H. Adelson
`
`Step la. Build Laplacian pyramids LA and LB for images A and B respectively.
`Step lb. Build a Gaussian pyramid GR for the region image R.
`Step 2. Form a combined pyramid LS from LA and LB using nodes of GR as
`weights. That is, for each l, i and j:
`
`LSl(i, j) = GRl(i, j)LAl(i, j) + (1 - GRl(i, j))LBl(i, j).
`
`Step 3. Obtain the splined image S by expanding and summing the levels of
`LS.
`
`The Gaussian pyramid serves two purposes here: It is a convenient method for
`determining which nodes at each pyramid level lie within the mask area of image
`R , and it "softens" the edges of the mask through an effective low-pass filter.
`Without this the spline would be overly sensitive to the position of the mask
`relative to the pyramid sample points. Nodes which fall exactly on the mask edge
`will receive a 50 percent weight, just as in the procedure outlined in the previous
`section. Here nodes to a distance of two sample positions on either side of the
`mask edge will also be combined as a weighted average of their LA and LB values
`(see Figure 5).
`An example using this technique is given in Figure 8. Figures 8a and 8b show
`the two original images, an eye and a hand. Figure 8c shows the region of the
`first image to be splined into the second image and Figure 8d shows the end
`result of the spline: a hand with an eye embedded in the palm.
`
`3.3 Splining Nonoverlapped Images
`
`Images must overlap if they are to be joined using any weighted average technique.
`Nonetheless, a satisfactory spline can be obtained with images that abut but do
`not overlap, if each image is first extrapolated across its boundary to form an
`overlapped transition zone. Since the width of the
`transition zone can be a
`significant fraction of the width of the image itself, extrapolation may at first
`seem to be a formidable task. However, in the multiresolution spline technique
`extrapolation can be performed separately in each frequency band. Furthermore,
`when the pyramid algorithm is used, only two samples need to be extrapolated
`beyond the edge of each level. In fact, it is just this type of extrapolation that is
`already handling boundary conditions during construction of both
`the Gaussian
`and Laplacian pyramids. No further steps need to be taken in the spline.
`An application of splining to nonoverlapping images is shown in Figure 9. We
`begin with a single image which is itself a mosaic of 16 b