[FFmpeg-devel] [PATCH] lavfi: edgedetect filter

Clément Bœsch ubitux at gmail.com
Tue Aug 7 23:57:15 CEST 2012


On Tue, Aug 07, 2012 at 08:44:25PM +0200, Clément Bœsch wrote:
[...]
> > > +static void gaussian_blur(AVFilterContext *ctx, int w, int h,
> > > +                                uint8_t *dst, int dst_linesize,
> > > +                          const uint8_t *src, int src_linesize)
> > > +{
> > > +    int i, j;
> > > +
> > > +    memcpy(dst,                src,                w);
> > > +    memcpy(dst + dst_linesize, src + src_linesize, w);
> > > +    for (j = 2; j < h - 2; j++) {
> > > +        dst[j*dst_linesize  ] = src[j*src_linesize  ];
> > > +        dst[j*dst_linesize+1] = src[j*src_linesize+1];
> > > +        for (i = 2; i < w - 2; i++) {
> > > +            dst[j*dst_linesize + i] =
> > > +                 ((src[(j-2)*src_linesize + (i-2)] + src[(j+2)*src_linesize + (i-2)]) * 2
> > > +                + (src[(j-2)*src_linesize + (i-1)] + src[(j+2)*src_linesize + (i-1)]) * 4
> > > +                + (src[(j-2)*src_linesize +  i   ] + src[(j+2)*src_linesize +  i   ]) * 5
> > > +                + (src[(j-2)*src_linesize + (i+1)] + src[(j+2)*src_linesize + (i+1)]) * 4
> > > +                + (src[(j-2)*src_linesize + (i+2)] + src[(j+2)*src_linesize + (i+2)]) * 2
> > > +
> > > +                + (src[(j-1)*src_linesize + (i-2)] + src[(j-1)*src_linesize + (i-2)]) *  4
> > > +                + (src[(j-1)*src_linesize + (i-1)] + src[(j-1)*src_linesize + (i-1)]) *  9
> > > +                + (src[(j-1)*src_linesize +  i   ] + src[(j-1)*src_linesize +  i   ]) * 12
> > > +                + (src[(j-1)*src_linesize + (i+1)] + src[(j-1)*src_linesize + (i+1)]) *  9
> > > +                + (src[(j-1)*src_linesize + (i+2)] + src[(j-1)*src_linesize + (i+2)]) *  4
> > > +
> > > +                + src[j*src_linesize + (i-2)] *  5
> > > +                + src[j*src_linesize + (i-1)] * 12
> > > +                + src[j*src_linesize +  i   ] * 15
> > > +                + src[j*src_linesize + (i+1)] * 12
> > > +                + src[j*src_linesize + (i+2)] *  5) / 159;
> > > +        }
> > > +        dst[j*dst_linesize + i  ] = src[j*src_linesize + i  ];
> > > +        dst[j*dst_linesize + i+1] = src[j*src_linesize + i+1];
> > > +    }
> > > +    memcpy(dst +  j   *dst_linesize, src +  j   *src_linesize, w);
> > > +    memcpy(dst + (j+1)*dst_linesize, src + (j+1)*src_linesize, w);
> > 
> > Is gcc smart enough to avoid all those multiplications by "linesize"? Just
> > adding "src += linesize; dst += linesize;" would probably be more readable
> > anyway.
> > 
> 
> I'll look at this deeper. I'm not sure doing increment all the time will
> really improve things...
> 

OK you were right, it's actually better. I did that in the other functions
as well.

Also it helped me to spot a shameful problem in that function (one column
of j-1 instead of j+1, now fixed and test ref updated).

> > > +}
> > > +
> > > +enum {
> > > +    DIRECTION_45UP,
> > > +    DIRECTION_45DOWN,
> > > +    DIRECTION_HORIZONTAL,
> > > +    DIRECTION_VERTICAL,
> > > +};
> > > +
> > > +static int get_rounded_direction(int gx, int gy)
> > > +{
> > > +    float tanpi8gx, tan3pi8gx;
> > > +
> > > +    /* reference angles:
> > > +     *   tan( pi/8) = sqrt(2)-1 ~= 0.41421...
> > > +     *   tan(3pi/8) = sqrt(2)+1 ~= 2.41421...
> > > +     * Gy/Gx is the tangent of theta, so Gy/Gx is compared against <ref-angle>,
> > > +     * or more simply Gy against <ref-angle>*Gx
> > > +     */
> > > +    if (gx) {
> > > +        if (gx < 0) // left side, switch signs
> > > +            gx = -gx, gy = -gy;
> > > +        tanpi8gx  = (M_SQRT2-1) * gx;
> > > +        tan3pi8gx = (M_SQRT2+1) * gx;
> > > +        if (gy > -tan3pi8gx && gy < -tanpi8gx)  return DIRECTION_45UP;
> > > +        if (gy > -tanpi8gx  && gy <  tanpi8gx)  return DIRECTION_HORIZONTAL;
> > > +        if (gy >  tanpi8gx  && gy <  tan3pi8gx) return DIRECTION_45DOWN;
> > > +    }
> > > +    return DIRECTION_VERTICAL;
> > > +}
> > 
> > A pure integer version would be nicer and more FATE-friendly. Since gx and
> > gy are bounded by ±256, the extra precision is not necessary:
> > 
> 
> mmh not really 256, more like [-1020;1020] afaict.
> 
> > if (gy << 16 > 158217) return ...
> > 
> > (16 instead of 8 if someone wants to implement higher bit depths)
> > 
> 
> I'll give it a try as well.
> 

Great it was a good idea. Tested with whole range of values, test
unchanged, and now bit exact. Thank you!

-- 
Clément B.
-------------- next part --------------
From 66705d9200c50f8f8fa7cc3874c7d494c5bd2c31 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Cl=C3=A9ment=20B=C5=93sch?= <ubitux at gmail.com>
Date: Thu, 26 Jul 2012 19:45:53 +0200
Subject: [PATCH] lavfi: edgedetect filter

FIXME: bump lavfi minor
---
 doc/filters.texi            |   5 +
 libavfilter/Makefile        |   1 +
 libavfilter/allfilters.c    |   1 +
 libavfilter/vf_edgedetect.c | 288 ++++++++++++++++++++++++++++++++++++++++++++
 tests/lavfi-regression.sh   |   1 +
 tests/ref/lavfi/edgedetect  |   1 +
 6 files changed, 297 insertions(+)
 create mode 100644 libavfilter/vf_edgedetect.c
 create mode 100644 tests/ref/lavfi/edgedetect

diff --git a/doc/filters.texi b/doc/filters.texi
index e73fc09..851d9bb 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -1929,6 +1929,11 @@ For more information about libfreetype, check:
 For more information about fontconfig, check:
 @url{http://freedesktop.org/software/fontconfig/fontconfig-user.html}.
 
+ at section edgedetect
+
+Detect and draw edges. The filter uses the Canny Edge Detection algorithm, with
+no parameter at the moment.
+
 @section fade
 
 Apply fade-in/out effect to input video.
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index b6bd37f..b8d89b0 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -89,6 +89,7 @@ OBJS-$(CONFIG_DELOGO_FILTER)                 += vf_delogo.o
 OBJS-$(CONFIG_DESHAKE_FILTER)                += vf_deshake.o
 OBJS-$(CONFIG_DRAWBOX_FILTER)                += vf_drawbox.o
 OBJS-$(CONFIG_DRAWTEXT_FILTER)               += vf_drawtext.o
+OBJS-$(CONFIG_EDGEDETECT_FILTER)             += vf_edgedetect.o
 OBJS-$(CONFIG_FADE_FILTER)                   += vf_fade.o
 OBJS-$(CONFIG_FIELDORDER_FILTER)             += vf_fieldorder.o
 OBJS-$(CONFIG_FIFO_FILTER)                   += fifo.o
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index da1c8e6..91eec2a 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -79,6 +79,7 @@ void avfilter_register_all(void)
     REGISTER_FILTER (DESHAKE,     deshake,     vf);
     REGISTER_FILTER (DRAWBOX,     drawbox,     vf);
     REGISTER_FILTER (DRAWTEXT,    drawtext,    vf);
+    REGISTER_FILTER (EDGEDETECT,  edgedetect,  vf);
     REGISTER_FILTER (FADE,        fade,        vf);
     REGISTER_FILTER (FIELDORDER,  fieldorder,  vf);
     REGISTER_FILTER (FIFO,        fifo,        vf);
diff --git a/libavfilter/vf_edgedetect.c b/libavfilter/vf_edgedetect.c
new file mode 100644
index 0000000..3299535
--- /dev/null
+++ b/libavfilter/vf_edgedetect.c
@@ -0,0 +1,288 @@
+/*
+ * Copyright (c) 2012 Clément Bœsch
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * Edge detection filter
+ *
+ * @url https://en.wikipedia.org/wiki/Canny_edge_detector
+ */
+
+#include "avfilter.h"
+#include "formats.h"
+#include "internal.h"
+#include "video.h"
+
+typedef struct {
+    uint8_t  *tmpbuf;
+    uint16_t *gradients;
+    char     *directions;
+} EdgeDetectContext;
+
+static int query_formats(AVFilterContext *ctx)
+{
+    static const enum PixelFormat pix_fmts[] = {PIX_FMT_GRAY8, PIX_FMT_NONE};
+    ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
+    return 0;
+}
+
+static int config_props(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx = inlink->dst;
+    EdgeDetectContext *edgedetect = ctx->priv;
+
+    edgedetect->tmpbuf     = av_malloc(inlink->w * inlink->h);
+    edgedetect->gradients  = av_calloc(inlink->w * inlink->h, sizeof(*edgedetect->gradients));
+    edgedetect->directions = av_malloc(inlink->w * inlink->h);
+    if (!edgedetect->tmpbuf || !edgedetect->gradients || !edgedetect->directions)
+        return AVERROR(ENOMEM);
+    return 0;
+}
+
+static void gaussian_blur(AVFilterContext *ctx, int w, int h,
+                                uint8_t *dst, int dst_linesize,
+                          const uint8_t *src, int src_linesize)
+{
+    int i, j;
+
+    memcpy(dst, src, w); dst += dst_linesize; src += src_linesize;
+    memcpy(dst, src, w); dst += dst_linesize; src += src_linesize;
+    for (j = 2; j < h - 2; j++) {
+        dst[0] = src[0];
+        dst[1] = src[1];
+        for (i = 2; i < w - 2; i++) {
+            dst[i] = ((src[-2*src_linesize + i-2] + src[2*src_linesize + i-2]) * 2
+                    + (src[-2*src_linesize + i-1] + src[2*src_linesize + i-1]) * 4
+                    + (src[-2*src_linesize + i  ] + src[2*src_linesize + i  ]) * 5
+                    + (src[-2*src_linesize + i+1] + src[2*src_linesize + i+1]) * 4
+                    + (src[-2*src_linesize + i+2] + src[2*src_linesize + i+2]) * 2
+
+                    + (src[  -src_linesize + i-2] + src[  src_linesize + i-2]) *  4
+                    + (src[  -src_linesize + i-1] + src[  src_linesize + i-1]) *  9
+                    + (src[  -src_linesize + i  ] + src[  src_linesize + i  ]) * 12
+                    + (src[  -src_linesize + i+1] + src[  src_linesize + i+1]) *  9
+                    + (src[  -src_linesize + i+2] + src[  src_linesize + i+2]) *  4
+
+                    + src[i-2] *  5
+                    + src[i-1] * 12
+                    + src[i  ] * 15
+                    + src[i+1] * 12
+                    + src[i+2] *  5) / 159;
+        }
+        dst[i    ] = src[i    ];
+        dst[i + 1] = src[i + 1];
+
+        dst += dst_linesize;
+        src += src_linesize;
+    }
+    memcpy(dst, src, w); dst += dst_linesize; src += src_linesize;
+    memcpy(dst, src, w); dst += dst_linesize; src += src_linesize;
+}
+
+enum {
+    DIRECTION_45UP,
+    DIRECTION_45DOWN,
+    DIRECTION_HORIZONTAL,
+    DIRECTION_VERTICAL,
+};
+
+static int get_rounded_direction(int gx, int gy)
+{
+    /* reference angles:
+     *   tan( pi/8) = sqrt(2)-1
+     *   tan(3pi/8) = sqrt(2)+1
+     * Gy/Gx is the tangent of the angle (theta), so Gy/Gx is compared against
+     * <ref-angle>, or more simply Gy against <ref-angle>*Gx
+     *
+     * Gx and Gy bounds = [1020;1020], using 16-bit arith:
+     *   round((sqrt(2)-1) * (1<<16)) =  27146
+     *   round((sqrt(2)+1) * (1<<16)) = 158218
+     */
+    if (gx) {
+        int tanpi8gx, tan3pi8gx;
+
+        if (gx < 0)
+            gx = -gx, gy = -gy;
+        gy <<= 16;
+        tanpi8gx  =  27145 * gx;
+        tan3pi8gx = 158218 * gx;
+        if (gy > -tan3pi8gx && gy < -tanpi8gx)  return DIRECTION_45UP;
+        if (gy > -tanpi8gx  && gy <  tanpi8gx)  return DIRECTION_HORIZONTAL;
+        if (gy >  tanpi8gx  && gy <  tan3pi8gx) return DIRECTION_45DOWN;
+    }
+    return DIRECTION_VERTICAL;
+}
+
+static void sobel(AVFilterContext *ctx, int w, int h,
+                        uint16_t *dst, int dst_linesize,
+                  const uint8_t  *src, int src_linesize)
+{
+    int i, j;
+    EdgeDetectContext *edgedetect = ctx->priv;
+
+    for (j = 1; j < h - 1; j++) {
+        dst += dst_linesize;
+        src += src_linesize;
+        for (i = 1; i < w - 1; i++) {
+            const int gx =
+                -1*src[-src_linesize + i-1] + 1*src[-src_linesize + i+1]
+                -2*src[                i-1] + 2*src[                i+1]
+                -1*src[ src_linesize + i-1] + 1*src[ src_linesize + i+1];
+            const int gy =
+                -1*src[-src_linesize + i-1] + 1*src[ src_linesize + i-1]
+                -2*src[-src_linesize + i  ] + 2*src[ src_linesize + i  ]
+                -1*src[-src_linesize + i+1] + 1*src[ src_linesize + i+1];
+
+            dst[i] = FFABS(gx) + FFABS(gy);
+            edgedetect->directions[j*w + i] = get_rounded_direction(gx, gy);
+        }
+    }
+}
+
+static void non_maximum_suppression(AVFilterContext *ctx, int w, int h,
+                                          uint8_t  *dst, int dst_linesize,
+                                    const uint16_t *src, int src_linesize)
+{
+    int i, j;
+    EdgeDetectContext *edgedetect = ctx->priv;
+
+#define COPY_MAXIMA(ay, ax, by, bx) do {                \
+    if (src[i] > src[(ay)*src_linesize + i+(ax)] &&     \
+        src[i] > src[(by)*src_linesize + i+(bx)])       \
+        dst[i] = av_clip_uint8(src[i]);                 \
+} while (0)
+
+    for (j = 1; j < h - 1; j++) {
+        dst += dst_linesize;
+        src += src_linesize;
+        for (i = 1; i < w - 1; i++) {
+            switch (edgedetect->directions[j*w + i]) {
+            case DIRECTION_45UP:        COPY_MAXIMA( 1, -1, -1,  1); break;
+            case DIRECTION_45DOWN:      COPY_MAXIMA(-1, -1,  1,  1); break;
+            case DIRECTION_HORIZONTAL:  COPY_MAXIMA( 0, -1,  0,  1); break;
+            case DIRECTION_VERTICAL:    COPY_MAXIMA(-1,  0,  1,  0); break;
+            }
+        }
+    }
+}
+
+static void double_threshold(AVFilterContext *ctx, int w, int h,
+                                   uint8_t *dst, int dst_linesize,
+                             const uint8_t *src, int src_linesize)
+{
+    int i, j;
+
+#define THRES_HIGH 80
+#define THRES_LOW  20
+
+    for (j = 0; j < h; j++) {
+        for (i = 0; i < w; i++) {
+            if (src[i] > THRES_HIGH) {
+                dst[i] = src[i];
+                continue;
+            }
+
+            if ((!i || i == w - 1 || !j || j == h - 1) &&
+                src[i] > THRES_LOW &&
+                (src[-src_linesize + i-1] > THRES_HIGH ||
+                 src[-src_linesize + i  ] > THRES_HIGH ||
+                 src[-src_linesize + i+1] > THRES_HIGH ||
+                 src[                i-1] > THRES_HIGH ||
+                 src[                i+1] > THRES_HIGH ||
+                 src[ src_linesize + i-1] > THRES_HIGH ||
+                 src[ src_linesize + i  ] > THRES_HIGH ||
+                 src[ src_linesize + i+1] > THRES_HIGH))
+                dst[i] = src[i];
+            else
+                dst[i] = 0;
+        }
+        dst += dst_linesize;
+        src += src_linesize;
+    }
+}
+
+static int end_frame(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx = inlink->dst;
+    EdgeDetectContext *edgedetect = ctx->priv;
+    AVFilterLink *outlink = inlink->dst->outputs[0];
+    AVFilterBufferRef  *inpicref = inlink->cur_buf;
+    AVFilterBufferRef *outpicref = outlink->out_buf;
+    uint8_t  *tmpbuf    = edgedetect->tmpbuf;
+    uint16_t *gradients = edgedetect->gradients;
+
+    /* gaussian filter to reduce noise  */
+    gaussian_blur(ctx, inlink->w, inlink->h,
+                  tmpbuf,            inlink->w,
+                  inpicref->data[0], inpicref->linesize[0]);
+
+    /* compute the 16-bits gradients and directions for the next step */
+    sobel(ctx, inlink->w, inlink->h,
+          gradients, inlink->w,
+          tmpbuf,    inlink->w);
+
+    /* non_maximum_suppression() will actually keep & clip what's necessary and
+     * ignore the rest, so we need a clean output buffer */
+    memset(tmpbuf, 0, inlink->w * inlink->h);
+    non_maximum_suppression(ctx, inlink->w, inlink->h,
+                            tmpbuf,    inlink->w,
+                            gradients, inlink->w);
+
+    /* keep high values, or low values surrounded by high values */
+    double_threshold(ctx, inlink->w, inlink->h,
+                     outpicref->data[0], outpicref->linesize[0],
+                     tmpbuf,             inlink->w);
+
+    ff_draw_slice(outlink, 0, outlink->h, 1);
+    return ff_end_frame(outlink);
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    EdgeDetectContext *edgedetect = ctx->priv;
+    av_freep(&edgedetect->tmpbuf);
+    av_freep(&edgedetect->gradients);
+    av_freep(&edgedetect->directions);
+}
+
+static int null_draw_slice(AVFilterLink *inlink, int y, int h, int slice_dir) { return 0; }
+
+AVFilter avfilter_vf_edgedetect = {
+    .name          = "edgedetect",
+    .description   = NULL_IF_CONFIG_SMALL("Detect and draw edge."),
+    .priv_size     = sizeof(EdgeDetectContext),
+    .uninit        = uninit,
+    .query_formats = query_formats,
+
+    .inputs    = (const AVFilterPad[]) {{ .name             = "default",
+                                          .type             = AVMEDIA_TYPE_VIDEO,
+                                          .draw_slice       = null_draw_slice,
+                                          .config_props     = config_props,
+                                          .end_frame        = end_frame,
+                                          .min_perms        = AV_PERM_READ
+                                        },
+                                        { .name = NULL }
+    },
+    .outputs   = (const AVFilterPad[]) {{ .name             = "default",
+                                          .type             = AVMEDIA_TYPE_VIDEO,
+                                        },
+                                        { .name = NULL }
+    },
+};
diff --git a/tests/lavfi-regression.sh b/tests/lavfi-regression.sh
index c763e99..0496d8a 100755
--- a/tests/lavfi-regression.sh
+++ b/tests/lavfi-regression.sh
@@ -42,6 +42,7 @@ do_lavfi "crop_scale"         "crop=iw-100:ih-100:100:100,scale=400:-1"
 do_lavfi "crop_scale_vflip"   "null,null,crop=iw-200:ih-200:200:200,crop=iw-20:ih-20:20:20,scale=200:200,scale=250:250,vflip,vflip,null,scale=200:200,crop=iw-100:ih-100:100:100,vflip,scale=200:200,null,vflip,crop=iw-100:ih-100:100:100,null"
 do_lavfi "crop_vflip"         "crop=iw-100:ih-100:100:100,vflip"
 do_lavfi "drawbox"            "drawbox=224:24:88:72:#FF8010 at 0.5"
+do_lavfi "edgedetect"         "edgedetect"
 do_lavfi "fade"               "fade=in:5:15,fade=out:30:15"
 do_lavfi "null"               "null"
 do_lavfi "overlay"            "split[m],scale=88:72,pad=96:80:4:4[o2];[m]fifo[o1],[o1][o2]overlay=240:16"
diff --git a/tests/ref/lavfi/edgedetect b/tests/ref/lavfi/edgedetect
new file mode 100644
index 0000000..52bae0b
--- /dev/null
+++ b/tests/ref/lavfi/edgedetect
@@ -0,0 +1 @@
+edgedetect          8fbed1b56636619f3b2d8d7b221c9242
-- 
1.7.11.4

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 490 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20120807/5d11325f/attachment.asc>


More information about the ffmpeg-devel mailing list