Fixes Ticket #4910
Clément Bœsch authored on 2015/10/01 03:29:30... | ... |
@@ -9695,6 +9695,41 @@ Negate input video. |
9695 | 9695 |
It accepts an integer in input; if non-zero it negates the |
9696 | 9696 |
alpha component (if available). The default value in input is 0. |
9697 | 9697 |
|
9698 |
+@section nlmeans |
|
9699 |
+ |
|
9700 |
+Denoise frames using Non-Local Means algorithm. |
|
9701 |
+ |
|
9702 |
+Each pixel is adjusted by looking for other pixels with similar contexts. This |
|
9703 |
+context similarity is defined by comparing their surrounding patches of size |
|
9704 |
+@option{p}x@option{p}. Patches are searched in an area of @option{r}x@option{r} |
|
9705 |
+around the pixel. |
|
9706 |
+ |
|
9707 |
+Note that the research area defines centers for patches, which means some |
|
9708 |
+patches will be made of pixels outside that research area. |
|
9709 |
+ |
|
9710 |
+The filter accepts the following options. |
|
9711 |
+ |
|
9712 |
+@table @option |
|
9713 |
+@item s |
|
9714 |
+Set denoising strength. |
|
9715 |
+ |
|
9716 |
+@item p |
|
9717 |
+Set patch size. |
|
9718 |
+ |
|
9719 |
+@item pc |
|
9720 |
+Same as @option{p} but for chroma planes. |
|
9721 |
+ |
|
9722 |
+The default value is @var{0} and means automatic. |
|
9723 |
+ |
|
9724 |
+@item r |
|
9725 |
+Set research size. |
|
9726 |
+ |
|
9727 |
+@item rc |
|
9728 |
+Same as @option{r} but for chroma planes. |
|
9729 |
+ |
|
9730 |
+The default value is @var{0} and means automatic. |
|
9731 |
+@end table |
|
9732 |
+ |
|
9698 | 9733 |
@section nnedi |
9699 | 9734 |
|
9700 | 9735 |
Deinterlace video using neural network edge directed interpolation. |
... | ... |
@@ -220,6 +220,7 @@ OBJS-$(CONFIG_METADATA_FILTER) += f_metadata.o |
220 | 220 |
OBJS-$(CONFIG_MINTERPOLATE_FILTER) += vf_minterpolate.o motion_estimation.o |
221 | 221 |
OBJS-$(CONFIG_MPDECIMATE_FILTER) += vf_mpdecimate.o |
222 | 222 |
OBJS-$(CONFIG_NEGATE_FILTER) += vf_lut.o |
223 |
+OBJS-$(CONFIG_NLMEANS_FILTER) += vf_nlmeans.o |
|
223 | 224 |
OBJS-$(CONFIG_NNEDI_FILTER) += vf_nnedi.o |
224 | 225 |
OBJS-$(CONFIG_NOFORMAT_FILTER) += vf_format.o |
225 | 226 |
OBJS-$(CONFIG_NOISE_FILTER) += vf_noise.o |
... | ... |
@@ -354,7 +355,7 @@ SKIPHEADERS-$(CONFIG_OPENCL) += opencl_internal.h deshake_opencl |
354 | 354 |
OBJS-$(CONFIG_SHARED) += log2_tab.o |
355 | 355 |
|
356 | 356 |
TOOLS = graph2dot |
357 |
-TESTPROGS = drawutils filtfmts formats |
|
357 |
+TESTPROGS = drawutils filtfmts formats integral |
|
358 | 358 |
|
359 | 359 |
TOOLS-$(CONFIG_LIBZMQ) += zmqsend |
360 | 360 |
|
... | ... |
@@ -237,6 +237,7 @@ void avfilter_register_all(void) |
237 | 237 |
REGISTER_FILTER(MINTERPOLATE, minterpolate, vf); |
238 | 238 |
REGISTER_FILTER(MPDECIMATE, mpdecimate, vf); |
239 | 239 |
REGISTER_FILTER(NEGATE, negate, vf); |
240 |
+ REGISTER_FILTER(NLMEANS, nlmeans, vf); |
|
240 | 241 |
REGISTER_FILTER(NNEDI, nnedi, vf); |
241 | 242 |
REGISTER_FILTER(NOFORMAT, noformat, vf); |
242 | 243 |
REGISTER_FILTER(NOISE, noise, vf); |
243 | 244 |
new file mode 100644 |
... | ... |
@@ -0,0 +1,92 @@ |
0 |
+/* |
|
1 |
+ * This file is part of FFmpeg. |
|
2 |
+ * |
|
3 |
+ * FFmpeg is free software; you can redistribute it and/or |
|
4 |
+ * modify it under the terms of the GNU Lesser General Public |
|
5 |
+ * License as published by the Free Software Foundation; either |
|
6 |
+ * version 2.1 of the License, or (at your option) any later version. |
|
7 |
+ * |
|
8 |
+ * FFmpeg is distributed in the hope that it will be useful, |
|
9 |
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of |
|
10 |
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
|
11 |
+ * Lesser General Public License for more details. |
|
12 |
+ * |
|
13 |
+ * You should have received a copy of the GNU Lesser General Public |
|
14 |
+ * License along with FFmpeg; if not, write to the Free Software |
|
15 |
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
|
16 |
+ */ |
|
17 |
+ |
|
18 |
+#include "libavfilter/vf_nlmeans.c" |
|
19 |
+ |
|
20 |
+static void display_integral(const uint32_t *ii, int w, int h, int lz_32) |
|
21 |
+{ |
|
22 |
+ int x, y; |
|
23 |
+ |
|
24 |
+ for (y = 0; y < h; y++) { |
|
25 |
+ for (x = 0; x < w; x++) |
|
26 |
+ printf(" %7x", ii[y*lz_32 + x]); |
|
27 |
+ printf("\n"); |
|
28 |
+ } |
|
29 |
+ printf("---------------\n"); |
|
30 |
+} |
|
31 |
+ |
|
32 |
+int main(void) |
|
33 |
+{ |
|
34 |
+ int ret = 0, xoff, yoff; |
|
35 |
+ |
|
36 |
+ // arbitrary test source of size 6x4 and linesize=8 |
|
37 |
+ const int w = 6, h = 5, lz = 8; |
|
38 |
+ static const uint8_t src[] = { |
|
39 |
+ 0xb0, 0x71, 0xfb, 0xd8, 0x01, 0xd9, /***/ 0x01, 0x02, |
|
40 |
+ 0x51, 0x8e, 0x41, 0x0f, 0x84, 0x58, /***/ 0x03, 0x04, |
|
41 |
+ 0xc7, 0x8d, 0x07, 0x70, 0x5c, 0x47, /***/ 0x05, 0x06, |
|
42 |
+ 0x09, 0x4e, 0xfc, 0x74, 0x8f, 0x9a, /***/ 0x07, 0x08, |
|
43 |
+ 0x60, 0x8e, 0x20, 0xaa, 0x95, 0x7d, /***/ 0x09, 0x0a, |
|
44 |
+ }; |
|
45 |
+ |
|
46 |
+ const int e = 3; |
|
47 |
+ const int ii_w = w+e*2, ii_h = h+e*2; |
|
48 |
+ |
|
49 |
+ // align to 4 the linesize, "+1" is for the space of the left 0-column |
|
50 |
+ const int ii_lz_32 = ((ii_w + 1) + 3) & ~3; |
|
51 |
+ |
|
52 |
+ // "+1" is for the space of the top 0-line |
|
53 |
+ uint32_t *ii = av_mallocz_array(ii_h + 1, ii_lz_32 * sizeof(*ii)); |
|
54 |
+ uint32_t *ii2 = av_mallocz_array(ii_h + 1, ii_lz_32 * sizeof(*ii2)); |
|
55 |
+ |
|
56 |
+ uint32_t *ii_start = ii + ii_lz_32 + 1; // skip top 0-line and left 0-column |
|
57 |
+ uint32_t *ii_start2 = ii2 + ii_lz_32 + 1; // skip top 0-line and left 0-column |
|
58 |
+ |
|
59 |
+ if (!ii || !ii2) |
|
60 |
+ return -1; |
|
61 |
+ |
|
62 |
+ for (yoff = -e; yoff <= e; yoff++) { |
|
63 |
+ for (xoff = -e; xoff <= e; xoff++) { |
|
64 |
+ int x, y; |
|
65 |
+ |
|
66 |
+ printf("xoff=%d yoff=%d\n", xoff, yoff); |
|
67 |
+ |
|
68 |
+ compute_ssd_integral_image(ii_start, ii_lz_32, |
|
69 |
+ src, lz, xoff, yoff, e, w, h); |
|
70 |
+ display_integral(ii_start, ii_w, ii_h, ii_lz_32); |
|
71 |
+ |
|
72 |
+ compute_unsafe_ssd_integral_image(ii_start2, ii_lz_32, |
|
73 |
+ 0, 0, |
|
74 |
+ src, lz, |
|
75 |
+ xoff, yoff, e, w, h, |
|
76 |
+ ii_w, ii_h); |
|
77 |
+ display_integral(ii_start2, ii_w, ii_h, ii_lz_32); |
|
78 |
+ |
|
79 |
+ if (memcmp(ii, ii2, (ii_h+1) * ii_lz_32 * sizeof(*ii))) { |
|
80 |
+ printf("Integral mismatch\n"); |
|
81 |
+ ret = 1; |
|
82 |
+ goto end; |
|
83 |
+ } |
|
84 |
+ } |
|
85 |
+ } |
|
86 |
+ |
|
87 |
+end: |
|
88 |
+ av_freep(&ii); |
|
89 |
+ av_freep(&ii2); |
|
90 |
+ return ret; |
|
91 |
+} |
... | ... |
@@ -30,7 +30,7 @@ |
30 | 30 |
#include "libavutil/version.h" |
31 | 31 |
|
32 | 32 |
#define LIBAVFILTER_VERSION_MAJOR 6 |
33 |
-#define LIBAVFILTER_VERSION_MINOR 62 |
|
33 |
+#define LIBAVFILTER_VERSION_MINOR 63 |
|
34 | 34 |
#define LIBAVFILTER_VERSION_MICRO 100 |
35 | 35 |
|
36 | 36 |
#define LIBAVFILTER_VERSION_INT AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \ |
37 | 37 |
new file mode 100644 |
... | ... |
@@ -0,0 +1,551 @@ |
0 |
+/* |
|
1 |
+ * Copyright (c) 2016 Clément Bœsch <u pkh me> |
|
2 |
+ * |
|
3 |
+ * This file is part of FFmpeg. |
|
4 |
+ * |
|
5 |
+ * FFmpeg is free software; you can redistribute it and/or |
|
6 |
+ * modify it under the terms of the GNU Lesser General Public |
|
7 |
+ * License as published by the Free Software Foundation; either |
|
8 |
+ * version 2.1 of the License, or (at your option) any later version. |
|
9 |
+ * |
|
10 |
+ * FFmpeg is distributed in the hope that it will be useful, |
|
11 |
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of |
|
12 |
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
|
13 |
+ * Lesser General Public License for more details. |
|
14 |
+ * |
|
15 |
+ * You should have received a copy of the GNU Lesser General Public |
|
16 |
+ * License along with FFmpeg; if not, write to the Free Software |
|
17 |
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
|
18 |
+ */ |
|
19 |
+ |
|
20 |
+/** |
|
21 |
+ * @todo |
|
22 |
+ * - SIMD for compute_safe_ssd_integral_image |
|
23 |
+ * - SIMD for final weighted averaging |
|
24 |
+ * - better automatic defaults? see "Parameters" @ http://www.ipol.im/pub/art/2011/bcm_nlm/ |
|
25 |
+ * - temporal support (probably doesn't need any displacement according to |
|
26 |
+ * "Denoising image sequences does not require motion estimation") |
|
27 |
+ * - Bayer pixel format support for at least raw photos? (DNG support would be |
|
28 |
+ * handy here) |
|
29 |
+ * - FATE test (probably needs visual threshold test mechanism due to the use |
|
30 |
+ * of floats) |
|
31 |
+ */ |
|
32 |
+ |
|
33 |
+#include "libavutil/avassert.h" |
|
34 |
+#include "libavutil/opt.h" |
|
35 |
+#include "libavutil/pixdesc.h" |
|
36 |
+#include "avfilter.h" |
|
37 |
+#include "formats.h" |
|
38 |
+#include "internal.h" |
|
39 |
+#include "video.h" |
|
40 |
+ |
|
41 |
+struct weighted_avg { |
|
42 |
+ double total_weight; |
|
43 |
+ double sum; |
|
44 |
+}; |
|
45 |
+ |
|
46 |
+#define WEIGHT_LUT_NBITS 9 |
|
47 |
+#define WEIGHT_LUT_SIZE (1<<WEIGHT_LUT_NBITS) |
|
48 |
+ |
|
49 |
+typedef struct { |
|
50 |
+ const AVClass *class; |
|
51 |
+ int nb_planes; |
|
52 |
+ int chroma_w, chroma_h; |
|
53 |
+ double pdiff_scale; // invert of the filtering parameter (sigma*10) squared |
|
54 |
+ double sigma; // denoising strength |
|
55 |
+ int patch_size, patch_hsize; // patch size and half size |
|
56 |
+ int patch_size_uv, patch_hsize_uv; // patch size and half size for chroma planes |
|
57 |
+ int research_size, research_hsize; // research size and half size |
|
58 |
+ int research_size_uv, research_hsize_uv; // research size and half size for chroma planes |
|
59 |
+ uint32_t *ii_orig; // integral image |
|
60 |
+ uint32_t *ii; // integral image starting after the 0-line and 0-column |
|
61 |
+ int ii_w, ii_h; // width and height of the integral image |
|
62 |
+ int ii_lz_32; // linesize in 32-bit units of the integral image |
|
63 |
+ struct weighted_avg *wa; // weighted average of every pixel |
|
64 |
+ int wa_linesize; // linesize for wa in struct size unit |
|
65 |
+ double weight_lut[WEIGHT_LUT_SIZE]; // lookup table mapping (scaled) patch differences to their associated weights |
|
66 |
+ double pdiff_lut_scale; // scale factor for patch differences before looking into the LUT |
|
67 |
+ int max_meaningful_diff; // maximum difference considered (if the patch difference is too high we ignore the pixel) |
|
68 |
+} NLMeansContext; |
|
69 |
+ |
|
70 |
+#define OFFSET(x) offsetof(NLMeansContext, x) |
|
71 |
+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM |
|
72 |
+static const AVOption nlmeans_options[] = { |
|
73 |
+ { "s", "denoising strength", OFFSET(sigma), AV_OPT_TYPE_DOUBLE, { .dbl = 1.0 }, 1.0, 30.0, FLAGS }, |
|
74 |
+ { "p", "patch size", OFFSET(patch_size), AV_OPT_TYPE_INT, { .i64 = 3*2+1 }, 0, 99, FLAGS }, |
|
75 |
+ { "pc", "patch size for chroma planes", OFFSET(patch_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 99, FLAGS }, |
|
76 |
+ { "r", "research window", OFFSET(research_size), AV_OPT_TYPE_INT, { .i64 = 7*2+1 }, 0, 99, FLAGS }, |
|
77 |
+ { "rc", "research window for chroma planes", OFFSET(research_size_uv), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 99, FLAGS }, |
|
78 |
+ { NULL } |
|
79 |
+}; |
|
80 |
+ |
|
81 |
+AVFILTER_DEFINE_CLASS(nlmeans); |
|
82 |
+ |
|
83 |
+static int query_formats(AVFilterContext *ctx) |
|
84 |
+{ |
|
85 |
+ static const enum AVPixelFormat pix_fmts[] = { |
|
86 |
+ AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, |
|
87 |
+ AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, |
|
88 |
+ AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV444P, |
|
89 |
+ AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P, |
|
90 |
+ AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P, |
|
91 |
+ AV_PIX_FMT_YUVJ411P, |
|
92 |
+ AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP, |
|
93 |
+ AV_PIX_FMT_NONE |
|
94 |
+ }; |
|
95 |
+ |
|
96 |
+ AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts); |
|
97 |
+ if (!fmts_list) |
|
98 |
+ return AVERROR(ENOMEM); |
|
99 |
+ return ff_set_common_formats(ctx, fmts_list); |
|
100 |
+} |
|
101 |
+ |
|
102 |
+/* |
|
103 |
+ * M is a discrete map where every entry contains the sum of all the entries |
|
104 |
+ * in the rectangle from the top-left origin of M to its coordinate. In the |
|
105 |
+ * following schema, "i" contains the sum of the whole map: |
|
106 |
+ * |
|
107 |
+ * M = +----------+-----------------+----+ |
|
108 |
+ * | | | | |
|
109 |
+ * | | | | |
|
110 |
+ * | a| b| c| |
|
111 |
+ * +----------+-----------------+----+ |
|
112 |
+ * | | | | |
|
113 |
+ * | | | | |
|
114 |
+ * | | X | | |
|
115 |
+ * | | | | |
|
116 |
+ * | d| e| f| |
|
117 |
+ * +----------+-----------------+----+ |
|
118 |
+ * | | | | |
|
119 |
+ * | g| h| i| |
|
120 |
+ * +----------+-----------------+----+ |
|
121 |
+ * |
|
122 |
+ * The sum of the X box can be calculated with: |
|
123 |
+ * X = e-d-b+a |
|
124 |
+ * |
|
125 |
+ * See https://en.wikipedia.org/wiki/Summed_area_table |
|
126 |
+ * |
|
127 |
+ * The compute*_ssd functions compute the integral image M where every entry |
|
128 |
+ * contains the sum of the squared difference of every corresponding pixels of |
|
129 |
+ * two input planes of the same size as M. |
|
130 |
+ */ |
|
131 |
+static inline int get_integral_patch_value(const uint32_t *ii, int ii_lz_32, int x, int y, int p) |
|
132 |
+{ |
|
133 |
+ const int e = ii[(y + p ) * ii_lz_32 + (x + p )]; |
|
134 |
+ const int d = ii[(y + p ) * ii_lz_32 + (x - p - 1)]; |
|
135 |
+ const int b = ii[(y - p - 1) * ii_lz_32 + (x + p )]; |
|
136 |
+ const int a = ii[(y - p - 1) * ii_lz_32 + (x - p - 1)]; |
|
137 |
+ return e - d - b + a; |
|
138 |
+} |
|
139 |
+ |
|
140 |
+/** |
|
141 |
+ * Compute squared difference of the safe area (the zone where s1 and s2 |
|
142 |
+ * overlap). It is likely the largest integral zone, so it is interesting to do |
|
143 |
+ * as little checks as possible; contrary to the unsafe version of this |
|
144 |
+ * function, we do not need any clipping here. |
|
145 |
+ * |
|
146 |
+ * The line above dst and the column to its left are always readable. |
|
147 |
+ * |
|
148 |
+ * This C version computes the SSD integral image using a scalar accumulator, |
|
149 |
+ * while for SIMD implementation it is likely more interesting to use the |
|
150 |
+ * two-loops algorithm variant. |
|
151 |
+ */ |
|
152 |
+static void compute_safe_ssd_integral_image_c(uint32_t *dst, int dst_linesize_32, |
|
153 |
+ const uint8_t *s1, int linesize1, |
|
154 |
+ const uint8_t *s2, int linesize2, |
|
155 |
+ int w, int h) |
|
156 |
+{ |
|
157 |
+ int x, y; |
|
158 |
+ |
|
159 |
+ for (y = 0; y < h; y++) { |
|
160 |
+ uint32_t acc = dst[-1] - dst[-dst_linesize_32 - 1]; |
|
161 |
+ |
|
162 |
+ for (x = 0; x < w; x++) { |
|
163 |
+ const int d = s1[x] - s2[x]; |
|
164 |
+ acc += d * d; |
|
165 |
+ dst[x] = dst[-dst_linesize_32 + x] + acc; |
|
166 |
+ } |
|
167 |
+ s1 += linesize1; |
|
168 |
+ s2 += linesize2; |
|
169 |
+ dst += dst_linesize_32; |
|
170 |
+ } |
|
171 |
+} |
|
172 |
+ |
|
173 |
+/** |
|
174 |
+ * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could |
|
175 |
+ * be readable). |
|
176 |
+ * |
|
177 |
+ * On the other hand, the line above dst and the column to its left are always |
|
178 |
+ * readable. |
|
179 |
+ * |
|
180 |
+ * There is little point in having this function SIMDified as it is likely too |
|
181 |
+ * complex and only handle small portions of the image. |
|
182 |
+ * |
|
183 |
+ * @param dst integral image |
|
184 |
+ * @param dst_linesize_32 integral image linesize (in 32-bit integers unit) |
|
185 |
+ * @param startx integral starting x position |
|
186 |
+ * @param starty integral starting y position |
|
187 |
+ * @param src source plane buffer |
|
188 |
+ * @param linesize source plane linesize |
|
189 |
+ * @param offx source offsetting in x |
|
190 |
+ * @param offy source offsetting in y |
|
191 |
+ * @paran r absolute maximum source offsetting |
|
192 |
+ * @param sw source width |
|
193 |
+ * @param sh source height |
|
194 |
+ * @param w width to compute |
|
195 |
+ * @param h height to compute |
|
196 |
+ */ |
|
197 |
+static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, int dst_linesize_32, |
|
198 |
+ int startx, int starty, |
|
199 |
+ const uint8_t *src, int linesize, |
|
200 |
+ int offx, int offy, int r, int sw, int sh, |
|
201 |
+ int w, int h) |
|
202 |
+{ |
|
203 |
+ int x, y; |
|
204 |
+ |
|
205 |
+ for (y = starty; y < starty + h; y++) { |
|
206 |
+ uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1]; |
|
207 |
+ const int s1y = av_clip(y - r, 0, sh - 1); |
|
208 |
+ const int s2y = av_clip(y - (r + offy), 0, sh - 1); |
|
209 |
+ |
|
210 |
+ for (x = startx; x < startx + w; x++) { |
|
211 |
+ const int s1x = av_clip(x - r, 0, sw - 1); |
|
212 |
+ const int s2x = av_clip(x - (r + offx), 0, sw - 1); |
|
213 |
+ const uint8_t v1 = src[s1y*linesize + s1x]; |
|
214 |
+ const uint8_t v2 = src[s2y*linesize + s2x]; |
|
215 |
+ const int d = v1 - v2; |
|
216 |
+ acc += d * d; |
|
217 |
+ dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc; |
|
218 |
+ } |
|
219 |
+ } |
|
220 |
+} |
|
221 |
+ |
|
222 |
+/* |
|
223 |
+ * Compute the sum of squared difference integral image |
|
224 |
+ * http://www.ipol.im/pub/art/2014/57/ |
|
225 |
+ * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis |
|
226 |
+ * |
|
227 |
+ * @param ii integral image of dimension (w+e*2) x (h+e*2) with |
|
228 |
+ * an additional zeroed top line and column already |
|
229 |
+ * "applied" to the pointer value |
|
230 |
+ * @param ii_linesize_32 integral image linesize (in 32-bit integers unit) |
|
231 |
+ * @param src source plane buffer |
|
232 |
+ * @param linesize source plane linesize |
|
233 |
+ * @param offx x-offsetting ranging in [-e;e] |
|
234 |
+ * @param offy y-offsetting ranging in [-e;e] |
|
235 |
+ * @param w source width |
|
236 |
+ * @param h source height |
|
237 |
+ * @param e research padding edge |
|
238 |
+ */ |
|
239 |
+static void compute_ssd_integral_image(uint32_t *ii, int ii_linesize_32, |
|
240 |
+ const uint8_t *src, int linesize, int offx, int offy, |
|
241 |
+ int e, int w, int h) |
|
242 |
+{ |
|
243 |
+ // ii has a surrounding padding of thickness "e" |
|
244 |
+ const int ii_w = w + e*2; |
|
245 |
+ const int ii_h = h + e*2; |
|
246 |
+ |
|
247 |
+ // we center the first source |
|
248 |
+ const int s1x = e; |
|
249 |
+ const int s1y = e; |
|
250 |
+ |
|
251 |
+ // 2nd source is the frame with offsetting |
|
252 |
+ const int s2x = e + offx; |
|
253 |
+ const int s2y = e + offy; |
|
254 |
+ |
|
255 |
+ // get the dimension of the overlapping rectangle where it is always safe |
|
256 |
+ // to compare the 2 sources pixels |
|
257 |
+ const int startx_safe = FFMAX(s1x, s2x); |
|
258 |
+ const int starty_safe = FFMAX(s1y, s2y); |
|
259 |
+ const int endx_safe = FFMIN(s1x + w, s2x + w); |
|
260 |
+ const int endy_safe = FFMIN(s1y + h, s2y + h); |
|
261 |
+ |
|
262 |
+ // top part where only one of s1 and s2 is still readable, or none at all |
|
263 |
+ compute_unsafe_ssd_integral_image(ii, ii_linesize_32, |
|
264 |
+ 0, 0, |
|
265 |
+ src, linesize, |
|
266 |
+ offx, offy, e, w, h, |
|
267 |
+ ii_w, starty_safe); |
|
268 |
+ |
|
269 |
+ // fill the left column integral required to compute the central |
|
270 |
+ // overlapping one |
|
271 |
+ compute_unsafe_ssd_integral_image(ii, ii_linesize_32, |
|
272 |
+ 0, starty_safe, |
|
273 |
+ src, linesize, |
|
274 |
+ offx, offy, e, w, h, |
|
275 |
+ startx_safe, endy_safe - starty_safe); |
|
276 |
+ |
|
277 |
+ // main and safe part of the integral |
|
278 |
+ av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w); |
|
279 |
+ av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h); |
|
280 |
+ av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w); |
|
281 |
+ av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h); |
|
282 |
+ compute_safe_ssd_integral_image_c(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32, |
|
283 |
+ src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize, |
|
284 |
+ src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize, |
|
285 |
+ endx_safe - startx_safe, endy_safe - starty_safe); |
|
286 |
+ |
|
287 |
+ // right part of the integral |
|
288 |
+ compute_unsafe_ssd_integral_image(ii, ii_linesize_32, |
|
289 |
+ endx_safe, starty_safe, |
|
290 |
+ src, linesize, |
|
291 |
+ offx, offy, e, w, h, |
|
292 |
+ ii_w - endx_safe, endy_safe - starty_safe); |
|
293 |
+ |
|
294 |
+ // bottom part where only one of s1 and s2 is still readable, or none at all |
|
295 |
+ compute_unsafe_ssd_integral_image(ii, ii_linesize_32, |
|
296 |
+ 0, endy_safe, |
|
297 |
+ src, linesize, |
|
298 |
+ offx, offy, e, w, h, |
|
299 |
+ ii_w, ii_h - endy_safe); |
|
300 |
+} |
|
301 |
+ |
|
302 |
+static int config_input(AVFilterLink *inlink) |
|
303 |
+{ |
|
304 |
+ AVFilterContext *ctx = inlink->dst; |
|
305 |
+ NLMeansContext *s = ctx->priv; |
|
306 |
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format); |
|
307 |
+ const int e = FFMAX(s->research_hsize, s->research_hsize_uv) |
|
308 |
+ + FFMAX(s->patch_hsize, s->patch_hsize_uv); |
|
309 |
+ |
|
310 |
+ s->chroma_w = FF_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w); |
|
311 |
+ s->chroma_h = FF_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h); |
|
312 |
+ s->nb_planes = av_pix_fmt_count_planes(inlink->format); |
|
313 |
+ |
|
314 |
+ /* Allocate the integral image with extra edges of thickness "e" |
|
315 |
+ * |
|
316 |
+ * +_+-------------------------------+ |
|
317 |
+ * |0|0000000000000000000000000000000| |
|
318 |
+ * +-x-------------------------------+ |
|
319 |
+ * |0|\ ^ | |
|
320 |
+ * |0| ii | e | |
|
321 |
+ * |0| v | |
|
322 |
+ * |0| +-----------------------+ | |
|
323 |
+ * |0| | | | |
|
324 |
+ * |0|<->| | | |
|
325 |
+ * |0| e | | | |
|
326 |
+ * |0| | | | |
|
327 |
+ * |0| +-----------------------+ | |
|
328 |
+ * |0| | |
|
329 |
+ * |0| | |
|
330 |
+ * |0| | |
|
331 |
+ * +-+-------------------------------+ |
|
332 |
+ */ |
|
333 |
+ s->ii_w = inlink->w + e*2; |
|
334 |
+ s->ii_h = inlink->h + e*2; |
|
335 |
+ |
|
336 |
+ // align to 4 the linesize, "+1" is for the space of the left 0-column |
|
337 |
+ s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4); |
|
338 |
+ |
|
339 |
+ // "+1" is for the space of the top 0-line |
|
340 |
+ s->ii_orig = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig)); |
|
341 |
+ if (!s->ii_orig) |
|
342 |
+ return AVERROR(ENOMEM); |
|
343 |
+ |
|
344 |
+ // skip top 0-line and left 0-column |
|
345 |
+ s->ii = s->ii_orig + s->ii_lz_32 + 1; |
|
346 |
+ |
|
347 |
+ // allocate weighted average for every pixel |
|
348 |
+ s->wa_linesize = inlink->w; |
|
349 |
+ s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa)); |
|
350 |
+ if (!s->wa) |
|
351 |
+ return AVERROR(ENOMEM); |
|
352 |
+ |
|
353 |
+ return 0; |
|
354 |
+} |
|
355 |
+ |
|
356 |
+struct thread_data { |
|
357 |
+ const uint8_t *src; |
|
358 |
+ int src_linesize; |
|
359 |
+ int startx, starty; |
|
360 |
+ int endx, endy; |
|
361 |
+ const uint32_t *ii_start; |
|
362 |
+ int p; |
|
363 |
+}; |
|
364 |
+ |
|
365 |
+static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) |
|
366 |
+{ |
|
367 |
+ int x, y; |
|
368 |
+ NLMeansContext *s = ctx->priv; |
|
369 |
+ const struct thread_data *td = arg; |
|
370 |
+ const uint8_t *src = td->src; |
|
371 |
+ const int src_linesize = td->src_linesize; |
|
372 |
+ const int process_h = td->endy - td->starty; |
|
373 |
+ const int slice_start = (process_h * jobnr ) / nb_jobs; |
|
374 |
+ const int slice_end = (process_h * (jobnr+1)) / nb_jobs; |
|
375 |
+ const int starty = td->starty + slice_start; |
|
376 |
+ const int endy = td->starty + slice_end; |
|
377 |
+ |
|
378 |
+ for (y = starty; y < endy; y++) { |
|
379 |
+ for (x = td->startx; x < td->endx; x++) { |
|
380 |
+ const int patch_diff_sq = get_integral_patch_value(td->ii_start, s->ii_lz_32, x, y, td->p); |
|
381 |
+ if (patch_diff_sq < s->max_meaningful_diff) { |
|
382 |
+ struct weighted_avg *wa = &s->wa[y*s->wa_linesize + x]; |
|
383 |
+ const int weight_lut_idx = patch_diff_sq * s->pdiff_lut_scale; |
|
384 |
+ const double weight = s->weight_lut[weight_lut_idx]; // exp(-patch_diff_sq * s->pdiff_scale) |
|
385 |
+ wa->total_weight += weight; |
|
386 |
+ wa->sum += weight * src[y*src_linesize + x]; |
|
387 |
+ } |
|
388 |
+ } |
|
389 |
+ } |
|
390 |
+ return 0; |
|
391 |
+} |
|
392 |
+ |
|
393 |
+static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r, |
|
394 |
+ uint8_t *dst, int dst_linesize, |
|
395 |
+ const uint8_t *src, int src_linesize) |
|
396 |
+{ |
|
397 |
+ int x, y; |
|
398 |
+ int offx, offy; |
|
399 |
+ NLMeansContext *s = ctx->priv; |
|
400 |
+ /* patches center points cover the whole research window so the patches |
|
401 |
+ * themselves overflow the research window */ |
|
402 |
+ const int e = r + p; |
|
403 |
+ /* focus an integral pointer on the centered image (s1) */ |
|
404 |
+ const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e; |
|
405 |
+ |
|
406 |
+ memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa)); |
|
407 |
+ |
|
408 |
+ for (offy = -r; offy <= r; offy++) { |
|
409 |
+ for (offx = -r; offx <= r; offx++) { |
|
410 |
+ if (offx || offy) { |
|
411 |
+ struct thread_data td = { |
|
412 |
+ .src = src + offy*src_linesize + offx, |
|
413 |
+ .src_linesize = src_linesize, |
|
414 |
+ .startx = FFMAX(0, -offx), |
|
415 |
+ .starty = FFMAX(0, -offy), |
|
416 |
+ .endx = FFMIN(w, w - offx), |
|
417 |
+ .endy = FFMIN(h, h - offy), |
|
418 |
+ .ii_start = centered_ii + offy*s->ii_lz_32 + offx, |
|
419 |
+ .p = p, |
|
420 |
+ }; |
|
421 |
+ |
|
422 |
+ compute_ssd_integral_image(s->ii, s->ii_lz_32, |
|
423 |
+ src, src_linesize, |
|
424 |
+ offx, offy, e, w, h); |
|
425 |
+ ctx->internal->execute(ctx, nlmeans_slice, &td, NULL, |
|
426 |
+ FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx))); |
|
427 |
+ } |
|
428 |
+ } |
|
429 |
+ } |
|
430 |
+ for (y = 0; y < h; y++) { |
|
431 |
+ for (x = 0; x < w; x++) { |
|
432 |
+ struct weighted_avg *wa = &s->wa[y*s->wa_linesize + x]; |
|
433 |
+ |
|
434 |
+ // Also weight the centered pixel |
|
435 |
+ wa->total_weight += 1.0; |
|
436 |
+ wa->sum += 1.0 * src[y*src_linesize + x]; |
|
437 |
+ |
|
438 |
+ dst[y*dst_linesize + x] = av_clip_uint8(wa->sum / wa->total_weight); |
|
439 |
+ } |
|
440 |
+ } |
|
441 |
+ return 0; |
|
442 |
+} |
|
443 |
+ |
|
444 |
+static int filter_frame(AVFilterLink *inlink, AVFrame *in) |
|
445 |
+{ |
|
446 |
+ int i; |
|
447 |
+ AVFilterContext *ctx = inlink->dst; |
|
448 |
+ NLMeansContext *s = ctx->priv; |
|
449 |
+ AVFilterLink *outlink = ctx->outputs[0]; |
|
450 |
+ |
|
451 |
+ AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h); |
|
452 |
+ if (!out) { |
|
453 |
+ av_frame_free(&in); |
|
454 |
+ return AVERROR(ENOMEM); |
|
455 |
+ } |
|
456 |
+ av_frame_copy_props(out, in); |
|
457 |
+ |
|
458 |
+ for (i = 0; i < s->nb_planes; i++) { |
|
459 |
+ const int w = i ? s->chroma_w : inlink->w; |
|
460 |
+ const int h = i ? s->chroma_h : inlink->h; |
|
461 |
+ const int p = i ? s->patch_hsize_uv : s->patch_hsize; |
|
462 |
+ const int r = i ? s->research_hsize_uv : s->research_hsize; |
|
463 |
+ nlmeans_plane(ctx, w, h, p, r, |
|
464 |
+ out->data[i], out->linesize[i], |
|
465 |
+ in->data[i], in->linesize[i]); |
|
466 |
+ } |
|
467 |
+ |
|
468 |
+ av_frame_free(&in); |
|
469 |
+ return ff_filter_frame(outlink, out); |
|
470 |
+} |
|
471 |
+ |
|
472 |
+#define CHECK_ODD_FIELD(field, name) do { \ |
|
473 |
+ if (!(s->field & 1)) { \ |
|
474 |
+ s->field |= 1; \ |
|
475 |
+ av_log(ctx, AV_LOG_WARNING, name " size must be odd, " \ |
|
476 |
+ "setting it to %d\n", s->field); \ |
|
477 |
+ } \ |
|
478 |
+} while (0) |
|
479 |
+ |
|
480 |
+static av_cold int init(AVFilterContext *ctx) |
|
481 |
+{ |
|
482 |
+ int i; |
|
483 |
+ NLMeansContext *s = ctx->priv; |
|
484 |
+ const double h = s->sigma * 10.; |
|
485 |
+ |
|
486 |
+ s->pdiff_scale = 1. / (h * h); |
|
487 |
+ s->max_meaningful_diff = -log(1/255.) / s->pdiff_scale; |
|
488 |
+ s->pdiff_lut_scale = 1./s->max_meaningful_diff * WEIGHT_LUT_SIZE; |
|
489 |
+ av_assert0((s->max_meaningful_diff - 1) * s->pdiff_lut_scale < FF_ARRAY_ELEMS(s->weight_lut)); |
|
490 |
+ for (i = 0; i < WEIGHT_LUT_SIZE; i++) |
|
491 |
+ s->weight_lut[i] = exp(-i / s->pdiff_lut_scale * s->pdiff_scale); |
|
492 |
+ |
|
493 |
+ CHECK_ODD_FIELD(research_size, "Luma research window"); |
|
494 |
+ CHECK_ODD_FIELD(patch_size, "Luma patch"); |
|
495 |
+ |
|
496 |
+ if (!s->research_size_uv) s->research_size_uv = s->research_size; |
|
497 |
+ if (!s->patch_size_uv) s->patch_size_uv = s->patch_size; |
|
498 |
+ |
|
499 |
+ CHECK_ODD_FIELD(research_size_uv, "Chroma research window"); |
|
500 |
+ CHECK_ODD_FIELD(patch_size_uv, "Chroma patch"); |
|
501 |
+ |
|
502 |
+ s->research_hsize = s->research_size / 2; |
|
503 |
+ s->research_hsize_uv = s->research_size_uv / 2; |
|
504 |
+ s->patch_hsize = s->patch_size / 2; |
|
505 |
+ s->patch_hsize_uv = s->patch_size_uv / 2; |
|
506 |
+ |
|
507 |
+ av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n", |
|
508 |
+ s->research_size, s->research_size, s->research_size_uv, s->research_size_uv, |
|
509 |
+ s->patch_size, s->patch_size, s->patch_size_uv, s->patch_size_uv); |
|
510 |
+ |
|
511 |
+ return 0; |
|
512 |
+} |
|
513 |
+ |
|
514 |
+static av_cold void uninit(AVFilterContext *ctx) |
|
515 |
+{ |
|
516 |
+ NLMeansContext *s = ctx->priv; |
|
517 |
+ av_freep(&s->ii_orig); |
|
518 |
+ av_freep(&s->wa); |
|
519 |
+} |
|
520 |
+ |
|
521 |
+static const AVFilterPad nlmeans_inputs[] = { |
|
522 |
+ { |
|
523 |
+ .name = "default", |
|
524 |
+ .type = AVMEDIA_TYPE_VIDEO, |
|
525 |
+ .config_props = config_input, |
|
526 |
+ .filter_frame = filter_frame, |
|
527 |
+ }, |
|
528 |
+ { NULL } |
|
529 |
+}; |
|
530 |
+ |
|
531 |
+static const AVFilterPad nlmeans_outputs[] = { |
|
532 |
+ { |
|
533 |
+ .name = "default", |
|
534 |
+ .type = AVMEDIA_TYPE_VIDEO, |
|
535 |
+ }, |
|
536 |
+ { NULL } |
|
537 |
+}; |
|
538 |
+ |
|
539 |
+AVFilter ff_vf_nlmeans = { |
|
540 |
+ .name = "nlmeans", |
|
541 |
+ .description = NULL_IF_CONFIG_SMALL("Non-local means denoiser."), |
|
542 |
+ .priv_size = sizeof(NLMeansContext), |
|
543 |
+ .init = init, |
|
544 |
+ .uninit = uninit, |
|
545 |
+ .query_formats = query_formats, |
|
546 |
+ .inputs = nlmeans_inputs, |
|
547 |
+ .outputs = nlmeans_outputs, |
|
548 |
+ .priv_class = &nlmeans_class, |
|
549 |
+ .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS, |
|
550 |
+}; |