Arcitecture for filtering images using Xilinx System Generator Alba M. Sáncez G., Ricardo Alvarez G., Sully Sáncez G.; FCC and FCE BUAP Abstract Tis paper presents an arcitecture for filters pixel by pixel and regions filters for image processing using Xilinx System Generator (XSG). Tis arcitecture offer an alternative troug a grapical user interface tat conbines MATLAB, Simulink and XSG and explore important aspects concerned to ardware implementation. Keywords Digital image processing, Matlab, Xilinx System Generator. T I. INTRODUCTION HE andling of digital images as become in recent decades a subject of widespread interest in different areas suc as medical and tecnological applications, among oters. We may cite lots of examples were image processing elps to analyze, infer and make decisions. Te main objective of image processing is to improve te quality of te images for uman interpretation, or te perception of te macines independently. Tis paper focuses in te processing pixel to pixel of an image and in te modification of pixel neigboroods and of course te transformation can be applied to te wole image or only a partial region. Te need to process te image in real time, leading to te implementation level ardware, wic offers parallelism, and tus significantly reduces te processing time, wic was wy decided to use Xilinx System Generator, a tool wit a grapical interface under te Matlab Simulink, based blocks wic makes it very easy to andle wit respect to oter software for ardware description. In addition to offering all te tools for an easy grapical simulation level. Te filtering of images is a tecnique wic gives an enancement of te image caracteristics. Te purpose of Manuscript received Marc 10, 2007; Revised May 31, 2007 Tis work was supported in part by te Puebla Benemerit Autonomus University (BUAP). Alba M. Sáncez G. is wit te Computer Science Faculty at te Puebla Benemerit Autonomus University, Av. San Claudio y 14 Sur Cd. Universitaria Z.P. 72570 (01 222 229 55 00 ext. 7218 e-mail: agalvez@ cs.buap.mx). Ricardo Alvarez G. is wit te Electronic Science Faculty at te Puebla Benemerit Autonomus University, Av. San Claudio y 18 Sur Cd. Universitaria Z.P. 72570 (01 222 229 55 00 ext. 7408 e-mail: algor@lece. buap.mx). Sully Sáncez G. is wit te Computer Science Faculty at te Puebla Benemerit Autonomus University, Av. San Claudio y 14 Sur Cd. Universitaria Z.P. 72570 (01 222 229 55 00 ext. 7217 e-mail: ssancez@ solarium.cs.buap.mx). Issue 2, Volume 1, 2007 101 filtering tecniques is to process an image in a way tat is more appropriate tan te original specific application. Among te applications of filter image are: te elimination of noise, enancing edges and contours, and so on. Tis article presents an arcitecture filtering images System Generator, wic is an extension of Simulink and consists of a bookstore called "blocks Xilinx", wic are mapped arcitectures, entities, signs, ports and attributes, wic Script file to produce syntesis in FPGAs, HDL simulation and developmentst tools. Te tool retains te ierarcy of Simulink wen it is converted into VHDL. II. XILINX SYSTEM GENERATOR Xilinx System Generator (XSG) is an integrated design environment (IDE) for FPGAs, wic uses Simulink, as a development environment and is presented in te form of blockset. It as an integrated design flow, to move directly to te configuration file (*. bit) necessary for programming te FPGA. One of te most important features of Xilinx System Generator is possessed abstraction aritmetic, tat is working wit representation in fixed point wit a precision arbitrary, including quantization and overflow. You can also perform simulations bot as a fixed point double precision. XSG automatically generates VHDL code and a draft of te ISE model being developed. Make ierarcical VHDL syntesis, expansion and mapping ardware, in addition to generating a user constraint file (UCF), simulation and testbec and test vectors among oter tings. Xilinx System Generator was created primarily to deal wit complex Digital Signal Processing (DSP) applications, but it as oter applications like te teme of tis work. Te blocks in Xilinx System Generator operate wit Boolean values or arbitrary values in fixed point, for a better approac to ardware implementation. In constrast Simulink works wit numbers of double-precision floating point. Te connection between blocks Xilinx System Generator and Simulink blocks are te gateway blocks. In Fig. 1 sows te broad flow design Xilinx System Generator. As already mentioned, you can ten move to te configuration file to program te FPGA, and te syntesis and implementation steps are optional for te user but not for
System Generator [5]. INTERNATIONAL JOURNAL of MATHEMATICS AND COMPUTERS IN SIMULATION Fig. 3. Block diagram of image processing. Fig. 1. Design flow in System Generator. III. DESIGN For te design of filters sould be meet ardware requirements, it is terefore necessary for image preprocessing prior to te same arcitecture. Unlike te level software processing, were te image is a two-dimensional arrangement n x m, and processed as suc, at ardware tis matrix must be an array of one dimension, namely a vector. Ten save te image information in a ROM memory. [6]. IV. IMAGE PROCESSING We are presenting te main filter processing tecnique of te picture: te processing pixel by pixel and te pixel neigborood [1]. Bot are working wit te intensity of te image. a) Processing pixel by pixel Tis tecnique is te simplest. It works wit te intensity of eac pixel [2], as sown in Figure 4. Fig. 4. f(x,y) is te intensity of unprocessing píxel, in te output we ave oter intensity, te coordinates are uncanged, only te intensity was modificated. Fig. 2. Store te image values in a ROM memory Te coordinate (i, j) suffers te following transformation ( i, j) ( j 1) m + i Coordinate latter is te position tat it occupies in te ROM. We ave sorted by columns, is te first to say tis in memory te first column and ten te second and so on until end wit te values, you can coose to be kept in rows. Once processed image, it applies a reverse transformation tat reversed te process of a settlement to a one-dimensional twodimensional. To access te information, you must address te ROM tat contains te value or values of te pixels tat are required for processing, depending on te filter to be used and tus obtain te new values of te image. Terefore requires two more blocks, te generator and address block processing operations. As sown in Fig. 3, te processing stage consists of tree blocks: te address generator, wic feeds te block ROM, it is te entire image information, tis in turn is used by te block Operational information processing. b) Processing of pixel neigboroods Basically consist in transform te value of a pixel in te position (i, j) taking into account te values of neigbors pixels. For example, if we consider a pixel neigborood and multiply wit a different weigt, by consider te values of te neigbors, te result of tis amount is te value of te new pixel of te image output in te same position (i, j). All tat remains is to define te values of te weigts, wic is usually defining a mask wit constant values. Te mask is actually a filter, so tat depending on its nature, and will be te end result. For example, if we define te mask next f f ( i 1, j 1) f ( i 1, j) f ( i 1, j + 1) f ( i, j 1) f ( i, j) f ( i, j + 1) ( ) ( ) ( ) i + 1, j 1 f i + 1, j f i + 1, j + 1 11 21 31 12 22 32 13 23 33 In (1) we ave te strengts of te region's image dimension equal to te mask, in tis case is 3 3, in subsection (2) is te (2) (1) Issue 2, Volume 1, 2007 102
mask, i.e. te weigts it will multiply eac of te intensities of te pixels. So te intensity of f (i, j) is replaced by: g ( i, j) = 11f ( i 1, j 1) + 21f ( i, j 1) + 31f ( i + 1, j 1) + 12f ( i 1, j) + f ( i j) + f ( i+ 1, j) + f ( i 1, j+ 1) + f ( i, j+ 1) + f ( i+ 1, j 1) INTERNATIONAL JOURNAL of MATHEMATICS AND COMPUTERS IN SIMULATION 22, 32 13 23 33 + V. RESULT S Among te image operations we ave te sine, contrast tresold applied to te image in Fig. 5. Fig. 8 Tresold Fig. 5 Original image Te images wit low contrast can be due to various causes, suc as poor ligting, lack of dynamic range in te sensor or even incorrect selection of te opening of te lens during te capture of te image. Fig. 6 sows a typical conversion used to enance te contrast and in Fig. 7 arcitecture. Fig. 9 Tresold arcitecture Fig. 6 Contrast Te negatives digitized images are useful in many applications, suc as medical imaging and representation in potograps of a monocrome screen wit films wit te idea of using te resulting negative slides as normal. Fig. 10 illustrates te use of tis simple transformation and arcitecture in Fig. 11. Fig. 10 Negative Fig. 7 Arcitecture for contrast Te operator tresold transforms an image of binary output from an image of gray scale, were te level of transition is given by te input parameter p1. Fig. 8 sows te implementation of tis transformation and Fig. 9 arcitecture. Fig. 11 Negative arcitecture Fig. 13 sows te negative image of te figure 12 and te fig. 14 arcitecture. Issue 2, Volume 1, 2007 103
Fig. 12 Original Image Fig. 13 Color negative Te filtering operations based its operations in te convolution of te image using te so-called core convolution. Tere are two types of filters: ig-pass and low-pass, wic in te context of te teory of signals supposed to miss te first ig frequency signal and te latter casualties. In te case of te pictures we refer to specific frequencies. Tis ig frequency is associated wit sudden canges in intensity intervals small space, i.e. edges, wile low frequencies refer to slow canges in te intensity. In Fig. 17 sows te effect of a low-pass filter and te result of a ig pass filter in Fig. 18. Fig. 17 Low pass Fig. 18 Hig pass Fig. 14 Color negative arcityecture Figure 15 and figure 16 sows te brigtness and arcitecture respectively applied to te image of te Figure 5. Te softening is precisely to mitigate or eliminate te igfrequency components (edges and noise). In te case of te edge effect is a mitigating tem and in te case of noise, te desired effect is te elimination of it. Te softening can be done by statistical tecniques: Aritmetic mean, weigted average, alf Gaussian. Te simplest model is te aritmetic mean, it considers an average of te pixels in an environment focused on a pixel (i, j). Fig. 20 sows te result wen applying to te image of te fig. 19. Fig. 15 Brigtness Fig. 19 Original image Fig. 20 Aritmetic mean It is possible model for oter processes suc as te average weigted average, it considers canging te weigt tat te central pixel located in (i, j) to an arbitrary value positive, ten wat we sould do is to preserve te status output te convolution matrix, we need to cange te weigt of pixels, ie te sum of te coefficients of te matrix is unity, so be sure not leave te range. Tis transformation is sown in fig. 21. Fig. 16 Brigtness Arcitecture In te operations of pixel neigborood noted for its functionality operations like softened edges and extraction. Te former is aimed at improving te quality of te image, wile te latter allow for te underlying edges. Issue 2, Volume 1, 2007 104
Fig. 21 Weigt mean for p=12 Te average Gaussian filter is based and consider a Gaussian curve of revolution around te center of a pixel matrix convolution. Te fig. 22 sows te effect. Fig. 24 G = G x + Gy Te operator of Prewitt is similar to Sobel, differing in te coefficients of masks. In Fig. 26 sows te results after applying tis filter to te image of te figure 25. Fig. 22 Gaussian mean Te first derivative is zero in all regions of constant intensity and as a constant value trougout te transition of intensity. Te second derivative, in contrast, is zero at all points except at te beginning and end of a transition intensity. Te first derivative is used to detect te presence of an edge, as well as te sign of te second derivative. Tey are different operators for extracting edges based on te first and second derivative. Te first derivative operators are based on te concept of vector gradient and led to te filters: gradient image, Sobel Prewit and Roberts. Te gradient of an image f (x, y) at a point (x, y) is defined as a two-dimensional array, wic consists of te derivative in te x direction and te direction and Daugter, wic is denoted respectively Gx, Gy. Te values and Gx Gy, can be implemented by convolution of te image masks figure known as operators Sobel. Te results are sown in Fig. 23, once it is applied Sobel operator to te image of te fig. 18. Fig. 25 Original image Fig. 26 Image process in Fig. 27 G x, and G = G x + Gy G y Te arcitecture for processing Prewitt is sown in fig. 28, in Fig. 29, te sum of te absolute values. a) b) Fig. 23 Image process in a)gx, b) G y Fig. 28 Prewitt Arcitecture Issue 2, Volume 1, 2007 105
Nort Sout Fig. 29 Add modules Te operator Roberts, unlike te previous two, it marks only te points edge, witout reporting tem guidance. It is a very simple operator wo works very well in binary images. Fig. 30 sows te transformation of te image to apply te filter. East Weast Norteast Nortwest a) b) Figure 30 Image process in a)d 1 and, b)d 2 Souteast Soutwest Fig. 33 Ligting in te eigt directions D = D 1 + D Figure 31 2 Witin te second derivative traders, igligts te Laplacian operator. In fig. 32 sows te result of applying te Laplace operator. VI. CONCLUSIONS Te Xilinx System Generator tool is a new application in image processing, and offers a friendly environment design for te processing, because te filters are designed by blocks. Tis tool support software simulation, but te most important is tat can syntesize in FPGAs ardware, wit te parallelism, robust and speed, tis features are essentials in image processing. REFERENCES Fig. 32 Laplacian In mapping, grapic arts and scientific applications are required to display information on te canges in te image, tis relates obviously directional derivative. Using anoter type of derivative can be acieved different ligting effects. Te fig. 33 was derived using te type Robinson wit te eigt directions given by te wind rose. Issue 2, Volume 1, 2007 106 [1] González R. C., Woods R. E.: Tratamiento digital de imágenes. Copublicación de Addison-Wesley Iberoamericana, S. A. y Ediciones Díaz de Santos, S. A. (1996) 180 243 [2] Pajares G., de la Cruz J. M., Molina J. M., Cuadrado J. López A.: Imágenes Digitales Procesamiento práctico con Java. Alfaomega, Ra-Ma (2004) 13-58 [3] Matlab website,ttp:// www.matworks.com. [4] Pajares G., de la Cruz J. M.: Visión por Computador Imágenes digitales y aplicaciones, Ra-Ma (2001) 89-177. [5] Xilinx System Generator User s Guide, ttp:// www. Xilinx.com. [6] Tejeda J.C., Arias M.: Arquitectura FPGA para filtrado de imágenes en Tiempo Real, Tesis de Maestría, INAOE,(2001).
[7] Notas de procesamiento de imágenes ttp://www.cs.buap.mx/~mmartin [8] R. J. Vidmar. (1992, August). On te use of atmosperic plasmas as electromagnetic reflectors. IEEE Trans. Plasma Sci. [Online]. 21(3). pp. 876 880. Available: ttp://www.alcyon.com/pub/journals/21ps03- vidmar Issue 2, Volume 1, 2007 107