libavfilter/vf_nlmeans.c

/*
 * Copyright (c) 2016 Clément Bœsch <u pkh me>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @todo
 * - SIMD for compute_safe_ssd_integral_image
 * - SIMD for final weighted averaging
 * - better automatic defaults? see "Parameters" @ http://www.ipol.im/pub/art/2011/bcm_nlm/
 * - temporal support (probably doesn't need any displacement according to
 *   "Denoising image sequences does not require motion estimation")
 * - Bayer pixel format support for at least raw photos? (DNG support would be
 *   handy here)
 * - FATE test (probably needs visual threshold test mechanism due to the use
 *   of floats)
 */

#include "libavutil/avassert.h"
#include "libavutil/opt.h"
#include "libavutil/pixdesc.h"
#include "avfilter.h"
#include "formats.h"
#include "internal.h"
#include "video.h"

struct weighted_avg {
    double total_weight;
    double sum;
};

#define WEIGHT_LUT_NBITS 9
#define WEIGHT_LUT_SIZE  (1<<WEIGHT_LUT_NBITS)

typedef struct NLMeansContext {
    const AVClass *class;
    int nb_planes;
    int chroma_w, chroma_h;
    double pdiff_scale;                         // invert of the filtering parameter (sigma*10) squared
    double sigma;                               // denoising strength
    int patch_size,    patch_hsize;             // patch size and half size
    int patch_size_uv, patch_hsize_uv;          // patch size and half size for chroma planes
    int research_size,    research_hsize;       // research size and half size
    int research_size_uv, research_hsize_uv;    // research size and half size for chroma planes
    uint32_t *ii_orig;                          // integral image
    uint32_t *ii;                               // integral image starting after the 0-line and 0-column
    int ii_w, ii_h;                             // width and height of the integral image
    int ii_lz_32;                               // linesize in 32-bit units of the integral image
    struct weighted_avg *wa;                    // weighted average of every pixel
    int wa_linesize;                            // linesize for wa in struct size unit
    double weight_lut[WEIGHT_LUT_SIZE];         // lookup table mapping (scaled) patch differences to their associated weights
    double pdiff_lut_scale;                     // scale factor for patch differences before looking into the LUT
    int max_meaningful_diff;                    // maximum difference considered (if the patch difference is too high we ignore the pixel)
} NLMeansContext;

#define OFFSET(x) offsetof(NLMeansContext, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
static const AVOption nlmeans_options[] = {
    { "s",  "denoising strength", OFFSET(sigma), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 }, 1.0, 30.0, FLAGS },
    { "p",  "patch size",                   OFFSET(patch_size),    AV_OPT_TYPE_INT, { .i64 = 3*2+1 }, 0, 99, FLAGS },
    { "pc", "patch size for chroma planes", OFFSET(patch_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 },     0, 99, FLAGS },
    { "r",  "research window",                   OFFSET(research_size),    AV_OPT_TYPE_INT, { .i64 = 7*2+1 }, 0, 99, FLAGS },
    { "rc", "research window for chroma planes", OFFSET(research_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 },     0, 99, FLAGS },
    { NULL }
};

AVFILTER_DEFINE_CLASS(nlmeans);

static int query_formats(AVFilterContext *ctx)
{
    static const enum AVPixelFormat pix_fmts[] = {
        AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
        AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
        AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P,
        AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
        AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
        AV_PIX_FMT_YUVJ411P,
        AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP,
        AV_PIX_FMT_NONE
    };

    AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
    if (!fmts_list)
        return AVERROR(ENOMEM);
    return ff_set_common_formats(ctx, fmts_list);
}

/*
 * M is a discrete map where every entry contains the sum of all the entries
 * in the rectangle from the top-left origin of M to its coordinate. In the
 * following schema, "i" contains the sum of the whole map:
 *
 * M = +----------+-----------------+----+
 *     |          |                 |    |
 *     |          |                 |    |
 *     |         a|                b|   c|
 *     +----------+-----------------+----+
 *     |          |                 |    |
 *     |          |                 |    |
 *     |          |        X        |    |
 *     |          |                 |    |
 *     |         d|                e|   f|
 *     +----------+-----------------+----+
 *     |          |                 |    |
 *     |         g|                h|   i|
 *     +----------+-----------------+----+
 *
 * The sum of the X box can be calculated with:
 *    X = e-d-b+a
 *
 * See https://en.wikipedia.org/wiki/Summed_area_table
 *
 * The compute*_ssd functions compute the integral image M where every entry
 * contains the sum of the squared difference of every corresponding pixels of
 * two input planes of the same size as M.
 */
static inline int get_integral_patch_value(const uint32_t *ii, int ii_lz_32, int x, int y, int p)
{
    const int e = ii[(y + p    ) * ii_lz_32 + (x + p    )];
    const int d = ii[(y + p    ) * ii_lz_32 + (x - p - 1)];
    const int b = ii[(y - p - 1) * ii_lz_32 + (x + p    )];
    const int a = ii[(y - p - 1) * ii_lz_32 + (x - p - 1)];
    return e - d - b + a;
}

/**
 * Compute squared difference of the safe area (the zone where s1 and s2
 * overlap). It is likely the largest integral zone, so it is interesting to do
 * as little checks as possible; contrary to the unsafe version of this
 * function, we do not need any clipping here.
 *
 * The line above dst and the column to its left are always readable.
 *
 * This C version computes the SSD integral image using a scalar accumulator,
 * while for SIMD implementation it is likely more interesting to use the
 * two-loops algorithm variant.
 */
static void compute_safe_ssd_integral_image_c(uint32_t *dst, int dst_linesize_32,
                                              const uint8_t *s1, int linesize1,
                                              const uint8_t *s2, int linesize2,
                                              int w, int h)
{
    int x, y;

    for (y = 0; y < h; y++) {
        uint32_t acc = dst[-1] - dst[-dst_linesize_32 - 1];

        for (x = 0; x < w; x++) {
            const int d  = s1[x] - s2[x];
            acc += d * d;
            dst[x] = dst[-dst_linesize_32 + x] + acc;
        }
        s1  += linesize1;
        s2  += linesize2;
        dst += dst_linesize_32;
    }
}

/**
 * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
 * be readable).
 *
 * On the other hand, the line above dst and the column to its left are always
 * readable.
 *
 * There is little point in having this function SIMDified as it is likely too
 * complex and only handle small portions of the image.
 *
 * @param dst               integral image
 * @param dst_linesize_32   integral image linesize (in 32-bit integers unit)
 * @param startx            integral starting x position
 * @param starty            integral starting y position
 * @param src               source plane buffer
 * @param linesize          source plane linesize
 * @param offx              source offsetting in x
 * @param offy              source offsetting in y
 * @paran r                 absolute maximum source offsetting
 * @param sw                source width
 * @param sh                source height
 * @param w                 width to compute
 * @param h                 height to compute
 */
static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, int dst_linesize_32,
                                                     int startx, int starty,
                                                     const uint8_t *src, int linesize,
                                                     int offx, int offy, int r, int sw, int sh,
                                                     int w, int h)
{
    int x, y;

    for (y = starty; y < starty + h; y++) {
        uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
        const int s1y = av_clip(y -  r,         0, sh - 1);
        const int s2y = av_clip(y - (r + offy), 0, sh - 1);

        for (x = startx; x < startx + w; x++) {
            const int s1x = av_clip(x -  r,         0, sw - 1);
            const int s2x = av_clip(x - (r + offx), 0, sw - 1);
            const uint8_t v1 = src[s1y*linesize + s1x];
            const uint8_t v2 = src[s2y*linesize + s2x];
            const int d = v1 - v2;
            acc += d * d;
            dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
        }
    }
}

/*
 * Compute the sum of squared difference integral image
 * http://www.ipol.im/pub/art/2014/57/
 * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
 *
 * @param ii                integral image of dimension (w+e*2) x (h+e*2) with
 *                          an additional zeroed top line and column already
 *                          "applied" to the pointer value
 * @param ii_linesize_32    integral image linesize (in 32-bit integers unit)
 * @param src               source plane buffer
 * @param linesize          source plane linesize
 * @param offx              x-offsetting ranging in [-e;e]
 * @param offy              y-offsetting ranging in [-e;e]
 * @param w                 source width
 * @param h                 source height
 * @param e                 research padding edge
 */
static void compute_ssd_integral_image(uint32_t *ii, int ii_linesize_32,
                                       const uint8_t *src, int linesize, int offx, int offy,
                                       int e, int w, int h)
{
    // ii has a surrounding padding of thickness "e"
    const int ii_w = w + e*2;
    const int ii_h = h + e*2;

    // we center the first source
    const int s1x = e;
    const int s1y = e;

    // 2nd source is the frame with offsetting
    const int s2x = e + offx;
    const int s2y = e + offy;

    // get the dimension of the overlapping rectangle where it is always safe
    // to compare the 2 sources pixels
    const int startx_safe = FFMAX(s1x, s2x);
    const int starty_safe = FFMAX(s1y, s2y);
    const int endx_safe   = FFMIN(s1x + w, s2x + w);
    const int endy_safe   = FFMIN(s1y + h, s2y + h);

    // top part where only one of s1 and s2 is still readable, or none at all
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      0, 0,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      ii_w, starty_safe);

    // fill the left column integral required to compute the central
    // overlapping one
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      0, starty_safe,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      startx_safe, endy_safe - starty_safe);

    // main and safe part of the integral
    av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
    av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
    av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
    av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
    compute_safe_ssd_integral_image_c(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
                                      src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize,
                                      src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize,
                                      endx_safe - startx_safe, endy_safe - starty_safe);

    // right part of the integral
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      endx_safe, starty_safe,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      ii_w - endx_safe, endy_safe - starty_safe);

    // bottom part where only one of s1 and s2 is still readable, or none at all
    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
                                      0, endy_safe,
                                      src, linesize,
                                      offx, offy, e, w, h,
                                      ii_w, ii_h - endy_safe);
}

static int config_input(AVFilterLink *inlink)
{
    AVFilterContext *ctx = inlink->dst;
    NLMeansContext *s = ctx->priv;
    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
    const int e = FFMAX(s->research_hsize, s->research_hsize_uv)
                + FFMAX(s->patch_hsize,    s->patch_hsize_uv);

    s->chroma_w = FF_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
    s->chroma_h = FF_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
    s->nb_planes = av_pix_fmt_count_planes(inlink->format);

    /* Allocate the integral image with extra edges of thickness "e"
     *
     *   +_+-------------------------------+
     *   |0|0000000000000000000000000000000|
     *   +-x-------------------------------+
     *   |0|\    ^                         |
     *   |0| ii  | e                       |
     *   |0|     v                         |
     *   |0|   +-----------------------+   |
     *   |0|   |                       |   |
     *   |0|<->|                       |   |
     *   |0| e |                       |   |
     *   |0|   |                       |   |
     *   |0|   +-----------------------+   |
     *   |0|                               |
     *   |0|                               |
     *   |0|                               |
     *   +-+-------------------------------+
     */
    s->ii_w = inlink->w + e*2;
    s->ii_h = inlink->h + e*2;

    // align to 4 the linesize, "+1" is for the space of the left 0-column
    s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);

    // "+1" is for the space of the top 0-line
    s->ii_orig = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig));
    if (!s->ii_orig)
        return AVERROR(ENOMEM);

    // skip top 0-line and left 0-column
    s->ii = s->ii_orig + s->ii_lz_32 + 1;

    // allocate weighted average for every pixel
    s->wa_linesize = inlink->w;
    s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
    if (!s->wa)
        return AVERROR(ENOMEM);

    return 0;
}

struct thread_data {
    const uint8_t *src;
    int src_linesize;
    int startx, starty;
    int endx, endy;
    const uint32_t *ii_start;
    int p;
};

static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
{
    int x, y;
    NLMeansContext *s = ctx->priv;
    const struct thread_data *td = arg;
    const uint8_t *src = td->src;
    const int src_linesize = td->src_linesize;
    const int process_h = td->endy - td->starty;
    const int slice_start = (process_h *  jobnr   ) / nb_jobs;
    const int slice_end   = (process_h * (jobnr+1)) / nb_jobs;
    const int starty = td->starty + slice_start;
    const int endy   = td->starty + slice_end;

    for (y = starty; y < endy; y++) {
        for (x = td->startx; x < td->endx; x++) {
            const int patch_diff_sq = get_integral_patch_value(td->ii_start, s->ii_lz_32, x, y, td->p);
            if (patch_diff_sq < s->max_meaningful_diff) {
                struct weighted_avg *wa = &s->wa[y*s->wa_linesize + x];
                const int weight_lut_idx = patch_diff_sq * s->pdiff_lut_scale;
                const double weight = s->weight_lut[weight_lut_idx]; // exp(-patch_diff_sq * s->pdiff_scale)
                wa->total_weight += weight;
                wa->sum += weight * src[y*src_linesize + x];
            }
        }
    }
    return 0;
}

static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r,
                         uint8_t *dst, int dst_linesize,
                         const uint8_t *src, int src_linesize)
{
    int x, y;
    int offx, offy;
    NLMeansContext *s = ctx->priv;
    /* patches center points cover the whole research window so the patches
     * themselves overflow the research window */
    const int e = r + p;
    /* focus an integral pointer on the centered image (s1) */
    const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e;

    memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));

    for (offy = -r; offy <= r; offy++) {
        for (offx = -r; offx <= r; offx++) {
            if (offx || offy) {
                struct thread_data td = {
                    .src          = src + offy*src_linesize + offx,
                    .src_linesize = src_linesize,
                    .startx       = FFMAX(0, -offx),
                    .starty       = FFMAX(0, -offy),
                    .endx         = FFMIN(w, w - offx),
                    .endy         = FFMIN(h, h - offy),
                    .ii_start     = centered_ii + offy*s->ii_lz_32 + offx,
                    .p            = p,
                };

                compute_ssd_integral_image(s->ii, s->ii_lz_32,
                                           src, src_linesize,
                                           offx, offy, e, w, h);
                ctx->internal->execute(ctx, nlmeans_slice, &td, NULL,
                                       FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
            }
        }
    }
    for (y = 0; y < h; y++) {
        for (x = 0; x < w; x++) {
            struct weighted_avg *wa = &s->wa[y*s->wa_linesize + x];

            // Also weight the centered pixel
            wa->total_weight += 1.0;
            wa->sum += 1.0 * src[y*src_linesize + x];

            dst[y*dst_linesize + x] = av_clip_uint8(wa->sum / wa->total_weight);
        }
    }
    return 0;
}

static int filter_frame(AVFilterLink *inlink, AVFrame *in)
{
    int i;
    AVFilterContext *ctx = inlink->dst;
    NLMeansContext *s = ctx->priv;
    AVFilterLink *outlink = ctx->outputs[0];

    AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
    if (!out) {
        av_frame_free(&in);
        return AVERROR(ENOMEM);
    }
    av_frame_copy_props(out, in);

    for (i = 0; i < s->nb_planes; i++) {
        const int w = i ? s->chroma_w          : inlink->w;
        const int h = i ? s->chroma_h          : inlink->h;
        const int p = i ? s->patch_hsize_uv    : s->patch_hsize;
        const int r = i ? s->research_hsize_uv : s->research_hsize;
        nlmeans_plane(ctx, w, h, p, r,
                      out->data[i], out->linesize[i],
                      in->data[i],  in->linesize[i]);
    }

    av_frame_free(&in);
    return ff_filter_frame(outlink, out);
}

#define CHECK_ODD_FIELD(field, name) do {                       \
    if (!(s->field & 1)) {                                      \
        s->field |= 1;                                          \
        av_log(ctx, AV_LOG_WARNING, name " size must be odd, "  \
               "setting it to %d\n", s->field);                 \
    }                                                           \
} while (0)

static av_cold int init(AVFilterContext *ctx)
{
    int i;
    NLMeansContext *s = ctx->priv;
    const double h = s->sigma * 10.;

    s->pdiff_scale = 1. / (h * h);
    s->max_meaningful_diff = -log(1/255.) / s->pdiff_scale;
    s->pdiff_lut_scale = 1./s->max_meaningful_diff * WEIGHT_LUT_SIZE;
    av_assert0((s->max_meaningful_diff - 1) * s->pdiff_lut_scale < FF_ARRAY_ELEMS(s->weight_lut));
    for (i = 0; i < WEIGHT_LUT_SIZE; i++)
        s->weight_lut[i] = exp(-i / s->pdiff_lut_scale * s->pdiff_scale);

    CHECK_ODD_FIELD(research_size,   "Luma research window");
    CHECK_ODD_FIELD(patch_size,      "Luma patch");

    if (!s->research_size_uv) s->research_size_uv = s->research_size;
    if (!s->patch_size_uv)    s->patch_size_uv    = s->patch_size;

    CHECK_ODD_FIELD(research_size_uv, "Chroma research window");
    CHECK_ODD_FIELD(patch_size_uv,    "Chroma patch");

    s->research_hsize    = s->research_size    / 2;
    s->research_hsize_uv = s->research_size_uv / 2;
    s->patch_hsize       = s->patch_size       / 2;
    s->patch_hsize_uv    = s->patch_size_uv    / 2;

    av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
           s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
           s->patch_size,    s->patch_size,    s->patch_size_uv,    s->patch_size_uv);

    return 0;
}

static av_cold void uninit(AVFilterContext *ctx)
{
    NLMeansContext *s = ctx->priv;
    av_freep(&s->ii_orig);
    av_freep(&s->wa);
}

static const AVFilterPad nlmeans_inputs[] = {
    {
        .name         = "default",
        .type         = AVMEDIA_TYPE_VIDEO,
        .config_props = config_input,
        .filter_frame = filter_frame,
    },
    { NULL }
};

static const AVFilterPad nlmeans_outputs[] = {
    {
        .name = "default",
        .type = AVMEDIA_TYPE_VIDEO,
    },
    { NULL }
};

AVFilter ff_vf_nlmeans = {
    .name          = "nlmeans",
    .description   = NULL_IF_CONFIG_SMALL("Non-local means denoiser."),
    .priv_size     = sizeof(NLMeansContext),
    .init          = init,
    .uninit        = uninit,
    .query_formats = query_formats,
    .inputs        = nlmeans_inputs,
    .outputs       = nlmeans_outputs,
    .priv_class    = &nlmeans_class,
    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,
};