Photometric Normalisation Algorithms

• Download demo project - 218 Kb

Output of multiscale retinex algorithm.

Introduction

The variation of illumination conditions of an object can produce large changes in the image plane, significantly impairing the performance of face verification and recognition algorithms. I present three photometric normalisation algorithms for use in pre-processing face images in order to be used in verification and recognition algorithms. Mainly I follow the ideas of paper "A Comparison of Photometric Normalisation Algorithms for Face Verification",James Short, Josef Kittler and Kieron Messer(2004) and "Lighting Normalization Algorithms for Face Verification",Guillaume Heusch ,Fabien Cardinaux, Sebastien Marcel(2005). Multiscale retinex method is coded exactly like the theory say. The anisotropic and isotropic smoothing methods have a little modifications but essentially its are the same. If you want to see more details about them you can see the papers previusly mentionated.

Using the code

You can apply the multiscale retinex method like:

        MultiscaleRetinex retinex = new MultiscaleRetinex(param.Sigmas, param.Widths, param.FilterSize);
        picFiltered.Image = retinex.Apply((Bitmap)bitmap.Clone());

This is the code of multiscale retinex algorithm:

        public override unsafe Bitmap Apply(Bitmap bitmap)
        {
            int count = sigmas.Length;
            Bitmap bmp;
            double[,] sum = new double;
            
            for (int i = 0; i < count;i++ )
            {
                bmp = new GaussianBlur(sigmas[i], size).Apply(bitmap);
                sum = SumBitmap(bmp,sum,widths[i]);
            }
            return Normalise(DivBitmap(bitmap, sum));
        }

You can apply the isotropic smoothing method like:

        IsotropicSmoothing iso = new IsotropicSmoothing(param.Value);
        picFiltered.Image = iso.Apply((Bitmap)bitmap.Clone());

This is the code of isotropic smoothing algorithm:

        public override unsafe Bitmap Apply(Bitmap bitmap)
        {
            Bitmap = bitmap;
            Point size = PixelSize;
            double[,] src = new double[size.X, size.Y];
            bool first = true;
            byte N, S, E, W, A;
            double Lw, Le, Ls, Ln, tmp, min = 0, max = 0;

            LockBitmap();

            for (int y = 0; y < size.Y ; y++)
            {
                PixelData* pPixel = PixelAt(0, y);
                for (int x = 0; x < size.X ; x++)
                {
                    tmp = pPixel->gray;

                    //Applying the process to all pixels of image except to the borders
                    if ((x > 0) && (x < size.X-1) && (y > 0) && (y < size.Y-1))
                    {
                        //Adjacent neighbouring pixels
                        A = pPixel->gray;           //current
                        E = PixelAt(x, y+1)->gray;  //east 
                        S = PixelAt(x+1, y)->gray;  //south 
                        N = PixelAt(x-1, y)->gray;  //north 
                        W = PixelAt(x, y-1)->gray;  //west  

                        //Ld refers to the derivative with respect to 
                        //each of the four adjacent neighbouring pixels
                        Lw = A - W;
                        Le = A - E;
                        Ln = A - N;
                        Ls = A - S;

                        //Isotropic smoothing
                        tmp = A + smooth * (Ln + Ls + Le + Lw);
                    }

                    src[x, y] = tmp;

                    //Computing the min and max values from all pixels of image
                    if (first) { min = max = tmp; first = false; }
                    else
                    {
                        if (tmp < min) min = tmp;
                        else
                            if (tmp > max) max = tmp;
                    }
                    pPixel++;
                }
            }
            UnlockBitmap();
            return Normalise(src, min, max);
        }

Output of isotropic smoothing algorithm.

You can apply the anisotropic smoothing method like:

        AnisotropicSmoothing anis = new AnisotropicSmoothing(param.Value);
        picFiltered.Image = anis.Apply((Bitmap)bitmap.Clone()); 

This is the code of anisotropic smoothing algorithm:

        public override unsafe Bitmap Apply(Bitmap bitmap) 
        {
            Bitmap = bitmap;
            Point size = PixelSize;
            double[,] src = new double[size.X,size.Y];
            bool first = true;
            byte N, S, E, W, A;
            double Lw, Le, Ls, Ln, pw, pe, ps, pn, eps = .1, tmp, min = 0, max = 0;
            
            LockBitmap();
            
            for (int y = 0; y < size.Y; y++)
            {
                PixelData* pPixel = PixelAt(0, y);
                for (int x = 0; x < size.X; x++)
                {
                    tmp = pPixel->gray;

                    //Applying the process to all pixels of image except to the borders
                    if ((x > 0) && (x < size.X-1) && (y > 0) && (y < size.Y-1))
                    {
                        //Adjacent neighbouring pixels
                        A = pPixel->gray;           //current
                        E = PixelAt(x, y+1)->gray;  //east 
                        S = PixelAt(x+1, y)->gray;  //south 
                        N = PixelAt(x-1, y)->gray;  //north 
                        W = PixelAt(x, y-1)->gray;  //west  

                        //Ld refers to the derivative with respect to 
                        //each of the four adjacent neighbouring pixels
                        Lw = A - W;
                        Le = A - E;
                        Ln = A - N;
                        Ls = A - S;

                        //Weber�s contrast inverse
                        pw = Math.Min(A, W) / (Math.Abs(A - W) + eps);
                        pe = Math.Min(A, E) / (Math.Abs(A - E) + eps);
                        pn = Math.Min(A, N) / (Math.Abs(A - N) + eps);
                        ps = Math.Min(A, S) / (Math.Abs(A - S) + eps);

                        //Anisotropic smoothing
                        tmp = A + smooth * (Ln * pn + Ls * ps + Le * pe + Lw * pw);
                    }

                    src[x, y] = tmp;

                    //Computing the min and max values from all pixels of image
                    if (first) { min = max = tmp; first = false; }
                    else
                    {
                        if (tmp < min) min = tmp;
                        else
                            if (tmp > max) max = tmp;
                    }
                    pPixel++;
                }
            }
            UnlockBitmap();
            return Normalise(src,min, max);
        }

Output of anisotropic smoothing algorithm.

Versions

1.0 14 Nov 2006

Credits

Dr. Eduardo Garea(Adviser)

Dr. Edel Garc�a(Adviser)

Andrew Kirillov's Image Processing Lab in C#

Points of interest

Image Processing in general (filtering,enhancement,denoisy etc.),C#, ASP.NET, Matlab, Java.

You must Sign In to use this message board.

FAQ    Noise Tolerance Very HighHighMediumLowVery Low   Layout Expand Posts & RepliesThread ViewNo JavascriptNo JS + Preview   Per page 102550

Msgs 1 to 9 of 9 (Total in Forum: 9) (Refresh) FirstPrevNext

Weird output Tolga Birdal 6:26 8 Nov '09   Hey, You know that the output of your anisotropic smoothing is weird, right? The algorithm (I don't know about the function) looks ok, but I advise to check the numerics. I don't think the output is correct, because it's not a diffusion at all. Tolga Sign In·View Thread·PermaLink Bit color without loose face details Ariston Darmayuda 8:26 22 Mar '07   How to filter face with bit color (black and white) without loose the face details (if face different, after filtered still different). Can this algorithm do that? Thanks. Sign In·View Thread·PermaLink Re: Bit color without loose face details Christian Graus 8:37 22 Mar '07   You can use dithering, but even then, you can't possible go from true color to 1 bit without losing some detail. Christian Graus - Microsoft MVP - C++ Metal Musings - Rex and my new metal blog "I am working on a project that will convert a FORTRAN code to corresponding C++ code.I am not aware of FORTRAN syntax" ( spotted in the C++/CLI forum ) Sign In·View Thread·PermaLink Re: Bit color without loose face details mosquets 10:51 22 Mar '07   The goal of these algorithms is to obtain a face image normalized in order to apply some recognize or verify algorithms. You can loose some details but you only must preserve the main features like eyes,mouth and nose and your distance proportion. It was think to gray scale images. I hope it should be useful for you. Marcel Sign In·View Thread·PermaLink 2.00/5 (1 vote) Re: Bit color without loose face details Ariston Darmayuda 4:42 25 Mar '07   Yes, that what I want. The normalized image only draw the border of eye, mouth, mouth, and face round, with white color, else with black. Like letter recognizer, the letter is black, background is white. If use grayscale, the color is not in bit format (not black and white). Is there any image classification method like PCA but simpler? Thanks. Sign In·View Thread·PermaLink Color Image japacheco 11:43 22 Jan '07   Can I Use the same algohritm on a color Image 24bppRGB ? Antonio Sign In·View Thread·PermaLink Re: Color Image mosquets 5:05 23 Jan '07   The algorithm was prepared to 8bpp because it is sufficient to apply after the recognition algorithms. I have not test the result with rgb but you can pass to my implementation the red image(8bpp) and get the result and after repeat the same with another two colors image and at the end create your color image with the three result. I hope it should be useful for you. Marcel Sign In·View Thread·PermaLink How can I enhance the original image by the output of Multi Scale Retinex? Brian Lau 20:54 3 Dec '06   as subject~ thanks~ Brian Lau Sign In·View Thread·PermaLink Re: How can I enhance the original image by the output of Multi Scale Retinex? mosquets 5:17 4 Dec '06   Can you formulate with more details what exactly do you want? In the two papers commented in my article is explained the theory about multi scale retinex. Mainly you must choose how many combination of images generated by low pass filtering the original image with Gaussians of varying widths. I don't know a general combination to do that but in my case I used 3 images and the results were good. I hope it will be useful for you. Mosquets Sign In·View Thread·PermaLink 3.33/5 (5 votes)

Last Visit: 4:51 23 Nov '09     Last Update: 4:51 23 Nov '09 1