libpipi/pipi/filter/blur.c

/*
 *  libpipi       Proper image processing implementation library
 *  Copyright (c) 2004-2008 Sam Hocevar <sam@zoy.org>
 *                All Rights Reserved
 *
 *  $Id$
 *
 *  This library is free software. It comes without any warranty, to
 *  the extent permitted by applicable law. You can redistribute it
 *  and/or modify it under the terms of the Do What The Fuck You Want
 *  To Public License, Version 2, as published by Sam Hocevar. See
 *  http://sam.zoy.org/wtfpl/COPYING for more details.
 */

/*
 * blur.c: blur functions
 */

#include "config.h"
#include "common.h"

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>

#include "pipi.h"
#include "pipi_internals.h"

/* Any standard deviation below this value will be rounded up, in order
 * to avoid ridiculously low values. exp(-1/(2*0.2*0.2)) is < 10^-5 so
 * there is little chance that any value below 0.2 will be useful. */
#define BLUR_EPSILON 0.2

pipi_image_t *pipi_gaussian_blur(pipi_image_t *src, float radius)
{
    return pipi_gaussian_blur_ext(src, radius, radius, 0.0, 0.0, 0.0);
}

pipi_image_t *pipi_gaussian_blur_ext(pipi_image_t *src, float rx, float ry,
                                     float angle, float dx, float dy)
{
    pipi_image_t *ret;
    double *kernel;
    double Kx, Ky, t = 0.0, sint, cost, bbx, bby;
    int i, j, krx, kry, m, n;

    if(rx < BLUR_EPSILON) rx = BLUR_EPSILON;
    if(ry < BLUR_EPSILON) ry = BLUR_EPSILON;

    sint = sin(angle * M_PI / 180.);
    cost = cos(angle * M_PI / 180.);

    /* Compute the final ellipse's bounding box */
    bbx = sqrt(rx * rx * cost * cost + ry * ry * sint * sint);
    bby = sqrt(ry * ry * cost * cost + rx * rx * sint * sint);

    /* FIXME: the kernel becomes far too big with large values of dx, because
     * we grow both left and right. Fix the growing direction. */
    krx = (int)(3. * bbx + .99999 + ceil(abs(dx)));
    m = 2 * krx + 1;
    Kx = -1. / (2. * rx * rx);

    kry = (int)(3. * bby + .99999 + ceil(abs(dy)));
    n = 2 * kry + 1;
    Ky = -1. / (2. * ry * ry);

    kernel = malloc(m * n * sizeof(double));

    for(j = -kry; j <= kry; j++)
    {
        for(i = -krx; i <= krx; i++)
        {
            /* FIXME: this level of interpolation sucks. We should
             * interpolate on the full NxN grid for better quality. */
            static double const samples[] =
            {
                 .0,  .0, 1,
                -.40, -.40, 0.8,
                -.30,  .0,  0.9,
                -.40,  .40, 0.8,
                 .0,   .30, 0.9,
                 .40,  .40, 0.8,
                 .30,  .0,  0.9,
                 .40, -.40, 0.8,
                 .0,  -.30, 0.9,
            };
            double u, v, ex, ey, d = 0.;
            unsigned int k;

            for(k = 0; k < sizeof(samples) / sizeof(*samples) / 3; k++)
            {
                u = ((double)i + samples[k * 3]) * cost
                     - ((double)j + samples[k * 3 + 1]) * sint + dx;
                v = ((double)i + samples[k * 3]) * sint
                     + ((double)j + samples[k * 3 + 1]) * cost + dy;
                ex = Kx * u * u;
                ey = Ky * v * v;
                d += samples[k * 3 + 2] * exp(ex + ey);

                /* Do not interpolate if this is a standard gaussian. */
                if(!dx && !dy && !angle)
                    break;
            }

            kernel[(j + kry) * m + i + krx] = d;
            t += d;
        }
    }

    for(j = 0; j < n; j++)
        for(i = 0; i < m; i++)
            kernel[j * m + i] /= t;

    ret = pipi_convolution(src, m, n, kernel);

    free(kernel);

    return ret;
}

pipi_image_t *pipi_box_blur(pipi_image_t *src, int size)
{
    return pipi_box_blur_ext(src, size, size);
}

/* FIXME: split this into templates for wrap-around and proper gray support */
pipi_image_t *pipi_box_blur_ext(pipi_image_t *src, int m, int n)
{
    pipi_image_t *dst;
    pipi_pixels_t *srcp, *dstp;
    float *srcdata, *dstdata;
    double *acc;
    int x, y, w, h, i, j, size, gray;

    w = src->w;
    h = src->h;
    size = (2 * m + 1) * (2 * n + 1);

    gray = (src->last_modified == PIPI_PIXELS_Y_F);

    srcp = gray ? pipi_getpixels(src, PIPI_PIXELS_Y_F)
                : pipi_getpixels(src, PIPI_PIXELS_RGBA_F);
    srcdata = (float *)srcp->pixels;

    dst = pipi_new(w, h);
    dstp = gray ? pipi_getpixels(dst, PIPI_PIXELS_Y_F)
                : pipi_getpixels(dst, PIPI_PIXELS_RGBA_F);
    dstdata = (float *)dstp->pixels;

    acc = malloc(w * (gray ? 1 : 4) * sizeof(double));

    /* Step 1: fill the accumulator */
    for(x = 0; x < w; x++)
    {
        if(gray)
        {
            double t = 0.;

            for(j = -n; j <= n; j++)
            {
                int j2 = (j < 0) ? h - 1 - ((-j - 1) % h) : j % h;
                t += srcdata[j2 * w + x];
            }

            acc[x] = t;
        }
        else
        {
            double r = 0., g = 0., b = 0., a = 0.;

            for(j = -n; j <= n; j++)
            {
                int j2 = (j < 0) ? h - 1 - ((-j - 1) % h) : j % h;
                r += srcdata[4 * (j2 * w + x)];
                g += srcdata[4 * (j2 * w + x) + 1];
                b += srcdata[4 * (j2 * w + x) + 2];
                a += srcdata[4 * (j2 * w + x) + 3];
            }

            acc[4 * x] = r;
            acc[4 * x + 1] = g;
            acc[4 * x + 2] = b;
            acc[4 * x + 3] = a;
        }
    }

    /* Step 2: blur the image, line by line */
    for(y = 0; y < h; y++)
    {
        double r = 0., g = 0., b = 0., a = 0.;
        double t = 0.;

        /* 2.1: compute the first pixel */
        if(gray)
        {
            for(i = -m; i <= m; i++)
            {
                int i2 = (i < 0) ? w - 1 - ((-i - 1) % w) : i % w;
                t += acc[i2];
            }
        }
        else
        {
            for(i = -m; i <= m; i++)
            {
                int i2 = (i < 0) ? w - 1 - ((-i - 1) % w) : i % w;
                r += acc[4 * i2];
                g += acc[4 * i2 + 1];
                b += acc[4 * i2 + 2];
                a += acc[4 * i2 + 3];
            }
        }

        /* 2.2: iterate on the whole line */
        for(x = 0; x < w; x++)
        {
            int u, u2, v, v2;

            if(gray)
            {
                dstdata[y * w + x] = t / size;
            }
            else
            {
                dstdata[4 * (y * w + x)] = r / size;
                dstdata[4 * (y * w + x) + 1] = g / size;
                dstdata[4 * (y * w + x) + 2] = b / size;
                dstdata[4 * (y * w + x) + 3] = a / size;
            }

            u = x - m;
            u2 = (u < 0) ? w - 1 - ((-u - 1) % w) : u % w;
            v = x + m + 1;
            v2 = (v < 0) ? w - 1 - ((-v - 1) % w) : v % w;
            if(gray)
            {
                t = t - acc[u2] + acc[v2];
            }
            else
            {
                r = r - acc[4 * u2] + acc[4 * v2];
                g = g - acc[4 * u2 + 1] + acc[4 * v2 + 1];
                b = b - acc[4 * u2 + 2] + acc[4 * v2 + 2];
                a = a - acc[4 * u2 + 3] + acc[4 * v2 + 3];
            }
        }

        /* 2.3: update the accumulator */
        for(x = 0; x < w; x++)
        {
            int u, u2, v, v2;

            u = y - n;
            u2 = (u < 0) ? w - 1 - ((-u - 1) % w) : u % w;
            v = y + n + 1;
            v2 = (v < 0) ? w - 1 - ((-v - 1) % w) : v % w;
            if(gray)
            {
                acc[x] += srcdata[v2 * w + x] - srcdata[u2 * w + x];
            }
            else
            {
                int uoff = 4 * (u2 * w + x);
                int voff = 4 * (v2 * w + x);

                acc[4 * x] += srcdata[voff] - srcdata[uoff];
                acc[4 * x + 1] += srcdata[voff + 1] - srcdata[uoff + 1];
                acc[4 * x + 2] += srcdata[voff + 2] - srcdata[uoff + 2];
                acc[4 * x + 3] += srcdata[voff + 3] - srcdata[uoff + 3];
            }
        }
    }

    free(acc);

    return dst;
}