Virtual restoration of vintage photographic prints affected by foxing and water blotches

Transcription

1 Journal of Electronic Imaging 14(4), (Oct Dec 2005) Virtual restoration of vintage photographic prints affected by foxing and water blotches Filippo Stanco Livio Tenze Giovanni Ramponi University of Trieste Dipartimento di Elettrotecnica, Elettronica e Informatica via A. Valerio 10, Trieste, Italy fstanco@dmi.unict.it Abstract. We propose a new algorithm to digitally restore vintage photographic prints affected by foxing and water blotches. It semiautomatically recovers the defects utilizing the features of the stains. The restoration process enhances the residual image information still present in the area. It is composed of three different steps: inpainting, additive-multiplicative (A-M) modeling, and interpolation SPIE and IS&T. DOI: / Introduction Since its origin, photography has been considered an important way to document reality. For this reason, its diffusion was fast and wide. In these first 150 years, the technologies, the materials of support, and the conservation techniques have changed significantly. Despite care in their conservation, the first photographic prints were based on fragile materials, and hence they are easily affected by bad environmental conditions. Frequently, the prints present cracks, scratches, holes, added stamps, or text. 1 Moreover, chemical reactions between the print and some microorganisms produce stains over the photographic prints, and humidity and water cause blotches that change the aspect of the picture. Restoration of these prints is necessary. The physical restoration of the print is a time-consuming and delicate operation, which requires skilled personnel. Often the original print is scanned or an image is acquired with digital photographic equipment, to make it possible to use it in virtual museums and on commercial Web sites. For these digital applications, we propose in this paper two methods for the virtual, semiautomatic restoration of the digital version of the print. Virtual restoration has been proposed in many cases, 1 12 and has proved to be convenient for various reasons. First, the processing is reversible and the original print is not affected, so that novel restoration techniques can be attempted without risks. Moreover, the restoration process can be at least partially automated, and this enables the treatment of a vast number of samples in a short time, without engaging dedicated personnel. Finally, the cost of the process is small, so that any professional photographer or small museum can afford it. Paper 04149R received Sep. 6, 2004; revised manuscript received May 6, 2005; accepted for publication May 17, 2005; published online Nov. 15, /2005/14 4 /043008/10/$ SPIE and IS&T. Two typical and very common defects are foxing and water blotches. The term foxing was used for the first time in the 18th century to indicate the scattered reddishbrown the color of a fox spots on the surface of paper in old books The same technical word was introduced in photography to refer to a similar chemical damage on the prints. Foxing is characterized by a dark-brown center and an area where the color is smoothed Fig. 1. In the center, all of the original information is covered by the stain, and hence it is considered as lacking. The area around the center, on the contrary, can include residual original information that should be enhanced. The causes of foxing are not completely understood; it probably depends on joined fungal activity and metal-induced degradation. The paper used in the oldest prints has microorganisms that can remain latent for decades awaiting conditions appropriate for growth. Moreover, airborne spores can attach to the paper, creating colonies of foxing. Another element that seems to accelerate foxing is the presence of iron in the paper. Indeed, if the relative humidity is below 50% and we use modern paper without iron, the foxing is strongly reduced. Another aggressive and, at the same time, frequent menace to photographic prints comes from water, which can permeate portions of the paper and produce very visible stains on the picture. The result is the so-called water blotch Fig. 2, which is often characterized by having a vaguely round shape, a color darker than the neighborhood due to the dust which is attracted in the paper texture, and an even darker border where the dust accumulates. An important peculiarity of this type of damage, which differentiates it from the many other defects that a print can show, is that the blotch does not completely destroy the content of the picture in the affected area: such a region is darker, but the image details are still at least partially visible. In this paper, we propose an algorithm to restore photographic prints affected by foxing and water blotches. The rest of the paper is organized as follows. Section 2 provides the structure of the restoration process. Section 3 describes in detail how to detect the defects. Section 4 reports the procedure used to restore the damages. Section 5 shows the experimental results and a conclusions section ends the paper. Journal of Electronic Imaging

2 Fig. 1 Examples of foxing. 2 Structure of the Restoration Process Techniques for effectively eliminating foxing and water blotches are not known in the open literature. The restoration method we propose is composed of two separate detection techniques, devoted to foxing and to blotches, respectively, and a restoration phase, in which the detected areas are processed. The restoration process is almost the same for both types of defects thanks to the fact that, as mentioned, in both cases the affected area or at least a portion of it still possesses some residual image information, which is exploited for the best results. We present here a compact description of the whole restoration procedure, while its components are detailed in the following sections. For clarity, a block scheme of the procedure is shown in Fig. 3. The input image I is a color image, represented by its R, G, and B components. 2.1 Foxing Detection Due to the characteristics of this type of defect, the detection of foxing is based on color. The detection process aims to define two different portions in each stain: the central one, opaque, in which the aggression has completely destroyed the original data, and a semitransparent periphery Fig. 2 Examples of water blotches. where some image details are still recognizable. To achieve this result, the input image is first converted to the YC b C r color space via a conventional transformation. 16 Then, the Journal of Electronic Imaging

3 Finally, it is important to perform the third and last restoration step: interpolation. Indeed, especially in homogeneous image zones, the contour of the processed area can be visible after the two previous steps. This is avoided by a straightforward linear interpolation, operating in the direction orthogonal to the outline of the defect area. The three restoration steps just described operate without requiring any user intervention. Fig. 3 Flow chart of the algorithm. histogram of the C r component is evaluated. Our experiments have shown that, in foxed images, such a histogram shows a readily identifiable shape. Image areas characterized by a proper, easily identifiable subset of C r chrominance levels are labeled as the opaque portions of the foxing stains. After this first phase, the semitransparent portion of each stain is searched for by expanding its core subject to suitable constraints in the local variation of the C r component second phase. Note that both phases in the detection process are completely automated. 2.2 Water Blotch Detection The detection process for this type of defect acts only on the luminance component Y of the image. It is obtained using a morphological segmentation 17 performed over a filtered image where the edges are enhanced. 18 The detection process is complicated by the possible presence of image details in the area. To help circumventing this problem, the user may actually be required to repeat the bootstrap procedure more than once. 2.3 Restoration The restoration process is performed over each RGB color plane. The process is composed of three different steps: inpainting, additive-multiplicative A-M modeling, and interpolation. The first step is devoted only to the opaque portion of the foxed areas, the remaining ones are used to restore areas affected by both types of aggression. The inpainting technique 2 6 propagates data from the uncorrupted portion of an image to the interior of the core of a foxing stain. Its basic aim is to replace the unrecoverable image data under the opaque foxing layer with values that show good continuity with respect to the luminance of the area exterior to the stain. After the inpainting has been performed, the key property of the remaining restoration algorithm must be its ability to exploit the image information still available in the damaged areas. For this purpose, we have selected the A-M model first introduced in Ref. 19 for the restoration of scratches in old films. The model parameters are derived based on the image values in an uncorrupted area surrounding the defect. 20 Then, the damaged pixels are substituted with the luminance values yielded by the model. Note that the same blotch or foxing area may contain different types of image details; in this case, a single A-M model is not sufficient and the area is split into different portions where the most effective parameters are selected. 3 Detection of the Defects 3.1 Foxing Phase I The first phase of foxing detection determines the locations of the spots. As mentioned, the foxing damage is related to the presence of iron; hence, the predominant color is always the same: red. Usually, vintage photographic prints are gray or sepia, the foxing stains have a color completely out of the color range of the image. Following this consideration, in the histogram of chrominance matrix C r related to the red the foxing pixels are represented by the smallest bins on the right tail. If the image is gray scale, the C r chrominance matrix related to red is null 128, while if it is sepia it presents a histogram in which most bins are located in the right half 128 to 256 of the possible range. We have observed that in presence of foxing artifacts, the typical histogram of C r has a tail on the right formed by a set of small bins having almost uniform amplitude that is not present on the left. Moreover, the peak of the histogram is located in the left portion of the set of nonzero bins of the histogram. The detection procedure searches for all the connected image parts represented by the bins on the right tail. If the histogram has the structure just presented, we search for all the bins representing the damaged pixels. We start from the right of the histogram, and we mark as foxing this bin B n. We denote with h B i the height of the bin B i. For i=n,n 1,...,2, if h B i h B i 1 Th f1, then bin B i 1 is marked as foxing; otherwise the procedure is stopped. The center of the last bin marked as foxing gives the value a, which is used to perform a thresholding over the matrix C r. The matrix I F1 is a map where the coordinates of foxed pixels are represented as a 0 value. This detection step could extract isolated points; they do not represent relevant damaged areas, and hence are expunged using a simple 3 3 median filter Phase II In this stage, we extend the previously detected areas by finding all the pixels where the original information is only partially affected by foxing. They are characterized by a lighter coloring than the center of the foxing and their position is near the reddish-brown spot. Therefore, we search them starting from the previous detection map I F1. The output of this phase is a new map I F2 that initially is equal to I F1.If is a foxing stain detected in phase I, we define the following sets of pixels: = x,y x,y is 8 connected to x,y, 1 Journal of Electronic Imaging

4 N xy = 8 connected neighborhood of x,y, and N xy = N xy. 3 For each pixel x,y we denote as C r N xy the average red chrominance value in N xy.if C r x,y C r N xy Th f2, 4 then the pixel is labeled as foxed and stored in I F2. This procedure is interactively repeated until there is at least one new pixel labeled as damaged in the previous iteration. If the chrominance changes in C r are uniform the detection phase can overestimate the number of damaged pixels. To avoid this situation, when a pixel is defined as damaged, we change its C r x,y value to the average value C r N xy. In other words, we propagate the border values of over the new pixels that are labeled as damaged. 3.2 Water Blotches In the scientific literature, specific algorithms do not exist for the detection of water blotches. There are methods for an analogous but different problem, the gap in artwork. Usually, a gap is characterized by approximately uniform gray levels. In Ref. 11, the detection system operates iteratively: it checks whether or not the gray level of a pixel differs significantly from the uniform range, and hence it determines if this pixel can be classified as belonging to the crack. In Ref. 12, the classification is performed, optimizing a function based on the luminance uniformity and on the magnitude of the gradient values in the region. More precisely, a pixel is deemed to belong to the damaged area if the new area obtained by adding this pixel to the damaged area increases the value of a suitable function. Despite the accuracy of the results in locating cracks or gaps in artwork, these algorithms do not provide good results with water blotches. We have found that the best detection results for the problem at hand are obtained by combining user interaction with segmentation algorithms. In this case, the users manually select one or more points in the damaged area, and the remaining region of the damage is automatically detected. The first specific detection method for water blotches was introduced in Ref. 21. The algorithm first performs an edge detection, and then, starting from the selected point, the algorithm analyzes the surrounding area using a 3 3 window. The samples within the window that do not belong to the border extend the detection map. This method requires a thresholding of the edge detection result. If the threshold is high, it could happen that the edges do not delimit a closed area, and the detection goes outside the water blotch. Conversely, if the threshold is conservative, the detection selects too restricted a portion of the blotch and multiple user intervention is necessary to complete the detection. Usually, since it is preferable to choose thresholds that yield closed areas, the detection is completed only with numerous user selections. The algorithm for water blotches detection used in this paper combines automatic segmentation algorithms with the user interaction. 22 The user selects one point inside the 2 blotch and all the pixels that belong to that region are added to the detection. This procedure is repeated until the selection is completed. The issue is to reduce the user interaction by increasing the performance of the segmentation. A segmentation process partitions an image into its constituent parts or objects. The number of regions obtained can be higher than the segmentation that a user would make basing on his or her own perceptive ability. Our algorithm reduces these undesired regions using as input for the segmentation a preprocessed image Y r, that is obtained by applying n x times a rational filter 18 RF over the luminance component Y of the image in the YC b C r color space. 16 The RF enhances the image attenuating small image variations while it preserves edges. Similarly to other nonlinear operators, 23,24 the RF modulates the coefficients of a linear lowpass filter to limit its action in presence of luminance changes. We have found that the RF is simpler and more effective than other techniques. Different versions of this operator can be devised; in the one we have selected, for each pixel Y i, j the output of the filter is obtained according to the relation Y i 1,j + Y i +1,j 2Y i, j Y r i, j = Y i, j + k Y i 1,j Y i +1,j 2 + A Y i, j 1 + Y i, j +1 2Y i, j + k Y i, j 1 Y i, j A, 5 where k and A are parameters that control the filter and take positive values. 18 The resulting image Y r tends to still have well-marked large edges, while values inside a region are made homogeneous; hence, it is the ideal input for a segmentation algorithm. We segment the image Y r using a sequence of morphological reconstructions. 17 More precisely we use an opening by reconstruction followed by a closing by reconstruction. Opening is an erosion followed by a dilation, while opening by reconstruction is an erosion followed by a morphological reconstruction. After opening by a disk of radius r, several connected components of the original image are removed, but the shape of some of the remaining ones is dramatically modified. After morphological reconstruction, 17 the original shape of the not totally removed particles is restored. Similar considerations can be chosen for the closing by reconstruction. Reconstructionbased opening and closing are more effective than standard opening and closing at removing blemishes without affecting the overall shapes of the objects. The image Y s obtained performing opening and closing by reconstruction over the image Y r is the segmented version of the original image Y. t If t is the number of regions R i inside Y r such that i=1 R i =, it is simple to assert that Y r = R i. t i=1 Now, the user selects a region that corresponds to the blotch. This region is stored in the map I B. If this region is not an exhaustive detection, the user selects other regions to add to the previous one and I B is updated. The RF reduces the insignificant details and it enables us to use a small radius r in the morphological operators. The lower is the radius r, the higher is the possibility of main- 6 Journal of Electronic Imaging

5 Table 1 Algorithm parameters. = x,y I F2 x,y =0 I B x,y =0 8 Parameter Value Min Max Th f1 0.1 Th f2 2 A 5 k 0.03 n r D W L Th Th n /Th 2 taining the small significant details. For example, if the image contains text, this may be heavily corrupted by the morphological operators with large radius, and the restoration becomes impossible. Moreover, note that the role of the opening by reconstruction is to preserve the shape of the object. Indeed, if a classical opening is performed instead of the opening by reconstruction, the following closing by reconstruction is not able to recover the shape of the object. 4 Restoration 4.1 Inpainting Inpainting algorithms 3 6 propagate both the gradient direction and the colors of a band surrounding the hole inside the hole to be filled in. These methods enable us to preserve the edges. We use an algorithm that propagates the colors but does not use the gradient direction. If is a single foxing blotch in the map I F1, we denote with all the pixels of that are in the border: = x,y p,q N xy : p,q. For each pixel x,y we extract the neighborhood N xy as defined in Eq. 2. Then, we assign to I 1 x,y the average of the pixels in N xy and not belonging to. The procedure ends when all the pixels in are considered. The new image I 1 presents more homogeneous foxing spots. They are still visible even if their saturation is reduced. This happens because in this step we have replaced the flawed regions using their outlines, which are partially affected by foxing. 4.2 Additive/Multiplicative Model The restoration algorithm presented in this paper is based on an additive-multiplicative model first introduced in Ref. 19 for the restoration of line scratches in old films. If 7 is the corrupted region to restore, as mentioned in Ref. 20, a suitable model to describe in the image I 1 can be the following: I 1 = J +, 9 where J is an ideal uncorrupted version of I 1, and and are the parameters to be estimated. If the variance and the mean respectively denoted as var and M operators are applied to Eq. 9, we obtain var I 1 = 2 var J 10 M I 1 = M J +. Equation 10 can be used to estimate the and values. However, the variance and the mean of the uncorrupted image J are unknown. To solve Eq. 9, we approximate J with I 1, where is an uncorrupted area around the blotch. Therefore, the approximate values of and, and, respectively, can be obtained by solving Eq. 10. Subsequently, each pixel in the corrupted regions can be corrected as follows: I 2 = I 1 /, 11 where I 2 is the restored image. After this step, the area inside the blotch is restored and it appears free of artifacts. The uncorrupted area with width W is automatically extracted. To avoid using pixels that are too close to the blotch, and hence are unreliable, is automatically shifted by D pixels away from the contour. This shift ensures that an erroneous detection of the border does not affect the accuracy of the final result. Sometimes, can be situated in areas with significant details. These areas could be the border of the print i.e., between the image and the support, or in areas where there is an edge. These are regions whose variance is too high. If var Th 1, the method divides into n different blocks to avoid the artifacts, with n user defined. Obviously, if the blotch is split in different regions, the area will also be split. Let min=min I and max =max I be the minimum and maximum gray values of the blotch. We use the parameter p= max min /n to compute the separation values v i =min+ip of the regions i, where i=0,...,n. Note that v 0 =min and v n =max. The centers of the regions i are C i =min+ i 1/2 p, with i =1,...,n. We compare each pixel value in with C i and put the pixel in the closest set. Similarly, the area is split in j with j=1,...,n. Now, we pair i with, j where i, j=1,...,n. The basic concept is that areas to be coupled have to be similar; and similar areas have close median values. Our procedure compares median I j with the values median I i, where i=1,...,n, and couples the area j with the area i whose medians have difference less than a threshold Th 2. If this condition is not satisfied, then we suppose that there is no similarity. In this case we do not use i and, j but the whole regions and. Journal of Electronic Imaging

6 Fig. 4 a Image in Fig. 1 a after restoration and b image in Fig. 1 b after restoration. 4.3 Interpolation The final output image I 3 is obtained by copying the corresponding values from I 2, for all points which do not belong to the damage border line. For the latter type of points, a linear 1-D interpolation is performed across the border line. It involves a vector of length 2L+1. First, we perform a simple interpolation and then we assign values at all the possible pixels not considered in the previous steps. More precisely, the luminance gradient is evaluated for each pixel P i in the contour. Then, an array P k :i L k i+l of 2L+1 pixels centered in P i is considered: the data are chosen along the gradient direction in P i so that L samples belong to the corrupted region and L belong to the uncorrupted one. If we denote with P start the first pixel of the array and with P end the last one, a conventional linear interpolation can be performed according to the distance between the pixels. If d P j = P j P end / P start P end, with j =1,...,2L+1, is the normalized distance of each pixel P j in the array from the P start position, the new intensity values are I 3 P j =d P j I 2 P start + 1 d P j I 2 P end. 12 Fig. 5 Images in a Fig. 2 a, b Fig. 2 b, and c Fig. 2 c after restoration. Journal of Electronic Imaging

7 Fig. 6 a Histogram of C r and b a detail of the histogram in a. Journal of Electronic Imaging

8 Due to the fact that the gradient orientation can be very different even for neighboring pixels, the interpolation process may skip some samples. To each pixel of this type, we assign a gray-level value corresponding to the average of its already interpolated neighbors in a 3 3 mask. Obviously, this interpolation does not take into account pixels that belong to another part of the stain and its related undamaged area. For example, if the blotch area a is associated with b, the interpolation will not use the pixels in i with i a i=1,...,n and in j with j b j=1,...,n. 5 Experimental Results This section shows some results obtained by applying the proposed algorithms to antique photographic prints. Note that the algorithms were tested with uncompressed and compressed images, in the tagged image file format TIFF and JPEG format respectively. In all cases the algorithms provided good results. The performances of the algorithms cannot be quantitatively estimated via mean squared error MSE or peak signal-to-noise ratio PSNR due to the fact that undamaged versions of the pictures do not exist. The test images are real scans of photographic prints affected by foxing and/or water blotches. Neither can images produced by restorers be considered as original images to be used in the comparison phase. Physical restorers are not able to remove foxing and water blotches from the image, they only stop the phenomenon. Virtual restorers use commercial software to replace the areas damaged with other ones subjectively determined. We submitted our experimental results to the Fratelli Alinari S.p.A. experts. Fratelli Alinari 25 is the world s oldest photo archive, and its experts are leaders in the photographic restoration field. They appreciated that the original texture and details inside the stains are maintained. They also appreciate the limited number of user interactions. The algorithm parameters that ensure good results for vintage images are reported in Table 1. The same table reports also the minimum and maximum values of the acceptable range for each parameter. The symbol = means that the parameter cannot be changed. This happens for the thresholds Th f1 and Th f2 in foxing detection, where small variations produce large differences in the results, and for the RF, where we use those proposed in Ref. 18. The values for the parameters D, W, and L are chosen by considering their nature: the distance D between and ensures that the area is out of the damage, even if a wrong detection is performed. Moreover, this area cannot be too far from the stain because the higher is the distance, the higher is the probability to include details not related to the defect. Similar considerations can be done for the width W of.a wide area could be heterogeneous, but a thin area might be not representative. The interpolation stage is the only one that does not preserve the details, but replace them with other values. For this reason, we create a thin interpolation area using a small value for the parameter L. The areas and are split in n different subareas; n can assume a wide range of values, even if very good results are obtained using n=2. Despite the fact that number of parameters is high, we Fig. 7 Position of foxing defects after a phase I and b phase II. stress that they have not been adjusted for each image. For the user, the algorithm is simple and semiautomatic. Figures 4 and 5 present some processed images. It is possible to notice that the algorithm enhances the information inside the damaged area and does not replace it with a texture or with new gray-level values different from the original ones. This behavior enables the restoration of large damaged areas while preserving details and without introducing artifacts. In particular, the image in Fig. 5 a has details that have been preserved while eliminating only the water drop effect. Figure 6 reports a typical histogram of C r in the case of foxing. As is possible to note in Fig. 6 b, there is a tail on the right of the histogram formed by a set of small bins. They represent the out-of-range red values, and the detection procedure searches for all the connected image parts represented by these bins. Figure 7 shows the output of the two foxing detection phases. It is possible to see that in Fig. 7 a near the black areas, there is a periphery where the details are still recognizable. These areas are added to the detection map during phase II Fig. 7 b. As expressed in Sec. 4, can be located in areas with significant details edges, border of prints, etc.. In these cases, we split into different blocks in order to avoid artifacts. Figure 8 b shows how and of Fig. 8 a are split if n=2: the corresponding areas have the same color. Note that the restoration phase does not process the entire image, but only the damaged areas. Moreover, the foxing detection algorithm does not require any selection by the user and operates automatically. Even if the water blotch detection method requires some interaction with the user, the achieved results are independent of the location of the starting point, provided that it is inside the stain. Journal of Electronic Imaging

9 barely visible artifacts, enhancing the residual information in the area. The efficacy with limited user intervention is enough to consider the algorithm a valid counterpart of manual restoration performed using commercial software restoration. Acknowledgments We thank F.lli Alinari s.p.a. for its support and useful hints, and for providing all the pictures used in our experiments. This work has been partially supported by a grant of the Regione Friuli Venezia Giulia, and is conducted in the framework of the SCHEMA NoE IST Fig. 8 a Damaged area Ω black and undamaged area Ω white ; b Ω and Ω split with the corresponding areas having the same color. 6 Conclusions An algorithm to remove foxing and water blotches defects from vintage photographic prints was proposed. The method preserves the residual information where available and replaces an irreversible damaged area with data obtained by considering the area around the defect. The algorithm works with very limited user intervention for both detection and restoration. Our experiments show the efficacy of the method: the stains are removed with References 1. F. Stanco, G. Ramponi, and A. de Polo, Towards the automated restoration of old photographic prints: a survey, in Proc. IEEE EUROCON 2003, pp A. Criminisi, P. Prez, and K. Toyama, Region filling and object removal by exemplar-based inpainting, IEEE Trans. Image Process. 13 9, C. Ballester, M. Bertalmio, V. Caselles, G. Sapiro, and J. Verdera, Filling-in by joint interpolation of vector fields and gray levels, IEEE Trans. Image Process. 10 8, M. Bertalmio, G. Sapiro, V. Caselles, and C. Ballester, Image inpainting, in Proc. SIGGRAPH 2000, pp J. M. S. Masnou, Level lines based disocclusion, in Proc. ICIP 98, pp M. Oliveira, B. Bowen, R. McKenna, and Y. Chang, Fast digital image inpainting, in Proc., Int. Conf. on Visualization, Imaging and Image Processing VIIP 2001, pp A. Kokaram, Motion Picture Restoration, Springer D. Tegolo and F. Isgro, A genetic algorithm for scratch removal in static images, in Proc. ICIAP 2001, pp L. Tenze, G. Ramponi, and S. Carrato, Robust detection and correction of blotches in old films using spatio-temporal information, Proc. SPIE 4667, M. Barni, F. Bartolini, and V. Cappellini, Image processing for virtual restoration of artworks, IEEE Multimedia 7 2, A. D. Rosa, A. Bonacchi, V. Cappellini, and M. Barni, Image segmentation and region filling for virtual restoration of art-works, in Proc. ICIP 2001, Vol. 1, pp T. Beckwith, W. Swanson, and T. Iiams, Deterioration of paper: the cause and effect of foxing, U. Calif. Publ. Bio. Sci. 1 13, A. Jain, Fundamentals of Digital Signal Processing, Prentice-Hall, Englewood Cliffs, NJ L. Vincent, Morphological grayscale reconstruction in image analysis: applications and efficient algorithms, IEEE Trans. Image Process. 2, Apr G. Ramponi, The rational filter for image smoothing, IEEE Signal Process. Lett. 3, Mar L. Tenze and G. Ramponi, Line scratch removal in vintage film based on an additive/multiplicative model, in Proc. IEEE Workshop on Nonlinear Signal and Image Processing F. Stanco, G. Ramponi, and L. Tenze, Removal of semi-transparent blotches in old photographic prints, in Proc. 5th COST 276 Workshop, pp F. Stanco, L. Tenze, and A. D. Rosa, An improved method for water blotches detection and restoration, in Proc., IEEE ISSPIT 2004, pp Rome F. Stanco and G. Ramponi, Detection of water blotches in antique documents, in Proc. 8th COST 276 Workshop, Trondheim, Norway May. 23. S. Guillon, P. Baylou, M. Najim, and N. Keskes, Adaptive nonlinear filters for 2D and 3D image enhancement, Signal Process. 67, S. Mitra, H. Li, I. Lin, and T. Yu, A new class of nonlinear for image enhancement, in Proc. IEEE Int. Conf. on Acoustics, Speech, and Signal Processing ICASSP-91, Vol. 4, pp Journal of Electronic Imaging

10 Filippo Stanco received his MSc summa cum laude and PhD degrees in computer science from the University of Catania, Italy, in 1999 and 2003, respectively. In , he was a contract researcher with the University of Trieste, Italy, where his work involved virtual restoration of cultural heritage. He is currently a contract researcher with the University of Catania, Italy. His research interests include digital restoration, artifact removal, filtering, superresolution, interpolation, and texture analysis. He is a reviewer for several leading international conferences and journals. Livio Tenze received his degree in electronic engineering in October 1998 and his PhD degree in 2002, from the University of Trieste, Italy. He joined the Engine Direction Department of the Ferrari Formula 1 Team, Maranello MO, in 2002, where his work involved 1-D signal processing applied to the F1 engine, analysis and development of the traction control system, and physical systems modeling. He then joined ENTEOS, Trieste, Italy, where he is currently a researcher in telecommunication systems. His research interests include image and video processing, multimedia applications, old motion picture restoration, image compression via JPEG 2000, and ASIC development using VHDL. Giovanni Ramponi received his degree in electronic engineering summa cum laude in He was a researcher, then an associate professor, and since 2000 he has been a full professor of electronics with the Department of Electronics of the University of Trieste. His research interests include nonlinear digital signal processing, enhancement and feature extraction in images and image sequences, and image compression. He is the coinventor on various pending international patents and has published more than 120 papers in international journals, conference proceedings, and book chapters. Professor Ramponi was an associate editor of the IEEE Signal Processing Letters and is currently an associate editor of the IEEE Transactions on Image Processing and of the SPIE Journal of Electronic Imaging. He chaired the Technical Programme of NSIP-03 and of Eusipco-96. He has been the local responsible for various scientific activities and contracts both of the European Union EU and nationally. He has also participated in other European and national research projects. Professor Ramponi contributes to several undergraduate and graduate courses on analog and digital electronics and on digital signal processing. Journal of Electronic Imaging