| e28b1fa6 |
/*
* Copyright (c) 2018 Paul B Mahol
*
* 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
*/
#include "libavutil/audio_fifo.h"
#include "libavutil/opt.h"
#include "avfilter.h"
#include "audio.h"
#include "formats.h"
typedef struct DeclickChannel {
double *auxiliary;
double *detection;
double *acoefficients;
double *acorrelation;
double *tmp;
double *interpolated;
double *matrix;
int matrix_size;
double *vector;
int vector_size;
double *y;
int y_size;
uint8_t *click;
int *index;
unsigned *histogram;
int histogram_size;
} DeclickChannel;
typedef struct AudioDeclickContext {
const AVClass *class;
double w;
double overlap;
double threshold;
double ar;
double burst;
int method;
int nb_hbins;
int is_declip;
int ar_order;
int nb_burst_samples;
int window_size;
int hop_size;
int overlap_skip;
AVFrame *in;
AVFrame *out;
AVFrame *buffer;
AVFrame *is;
DeclickChannel *chan;
int64_t pts;
int nb_channels;
uint64_t nb_samples;
uint64_t detected_errors;
int samples_left;
AVAudioFifo *fifo;
double *window_func_lut;
int (*detector)(struct AudioDeclickContext *s, DeclickChannel *c,
double sigmae, double *detection,
double *acoefficients, uint8_t *click, int *index,
const double *src, double *dst);
} AudioDeclickContext;
#define OFFSET(x) offsetof(AudioDeclickContext, x)
#define AF AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
static const AVOption adeclick_options[] = {
{ "w", "set window size", OFFSET(w), AV_OPT_TYPE_DOUBLE, {.dbl=55}, 10, 100, AF },
{ "o", "set window overlap", OFFSET(overlap), AV_OPT_TYPE_DOUBLE, {.dbl=75}, 50, 95, AF },
{ "a", "set autoregression order", OFFSET(ar), AV_OPT_TYPE_DOUBLE, {.dbl=2}, 0, 25, AF },
{ "t", "set threshold", OFFSET(threshold), AV_OPT_TYPE_DOUBLE, {.dbl=2}, 1, 100, AF },
{ "b", "set burst fusion", OFFSET(burst), AV_OPT_TYPE_DOUBLE, {.dbl=2}, 0, 10, AF },
{ "m", "set overlap method", OFFSET(method), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, AF, "m" },
{ "a", "overlap-add", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, AF, "m" },
{ "s", "overlap-save", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, AF, "m" },
{ NULL }
};
AVFILTER_DEFINE_CLASS(adeclick);
static int query_formats(AVFilterContext *ctx)
{
AVFilterFormats *formats = NULL;
AVFilterChannelLayouts *layouts = NULL;
static const enum AVSampleFormat sample_fmts[] = {
AV_SAMPLE_FMT_DBLP,
AV_SAMPLE_FMT_NONE
};
int ret;
formats = ff_make_format_list(sample_fmts);
if (!formats)
return AVERROR(ENOMEM);
ret = ff_set_common_formats(ctx, formats);
if (ret < 0)
return ret;
layouts = ff_all_channel_counts();
if (!layouts)
return AVERROR(ENOMEM);
ret = ff_set_common_channel_layouts(ctx, layouts);
if (ret < 0)
return ret;
formats = ff_all_samplerates();
return ff_set_common_samplerates(ctx, formats);
}
static int config_input(AVFilterLink *inlink)
{
AVFilterContext *ctx = inlink->dst;
AudioDeclickContext *s = ctx->priv;
int i;
s->pts = AV_NOPTS_VALUE;
s->window_size = inlink->sample_rate * s->w / 1000.;
if (s->window_size < 100)
return AVERROR(EINVAL);
s->ar_order = FFMAX(s->window_size * s->ar / 100., 1);
s->nb_burst_samples = s->window_size * s->burst / 1000.;
s->hop_size = s->window_size * (1. - (s->overlap / 100.));
if (s->hop_size < 1)
return AVERROR(EINVAL);
s->window_func_lut = av_calloc(s->window_size, sizeof(*s->window_func_lut));
if (!s->window_func_lut)
return AVERROR(ENOMEM);
for (i = 0; i < s->window_size; i++)
s->window_func_lut[i] = sin(M_PI * i / s->window_size) *
(1. - (s->overlap / 100.)) * M_PI_2;
av_frame_free(&s->in);
av_frame_free(&s->out);
av_frame_free(&s->buffer);
av_frame_free(&s->is);
s->in = ff_get_audio_buffer(inlink, s->window_size);
s->out = ff_get_audio_buffer(inlink, s->window_size);
s->buffer = ff_get_audio_buffer(inlink, s->window_size * 2);
s->is = ff_get_audio_buffer(inlink, s->window_size);
if (!s->in || !s->out || !s->buffer || !s->is)
return AVERROR(ENOMEM);
s->fifo = av_audio_fifo_alloc(inlink->format, inlink->channels, s->window_size);
if (!s->fifo)
return AVERROR(ENOMEM);
s->overlap_skip = s->method ? (s->window_size - s->hop_size) / 2 : 0;
if (s->overlap_skip > 0) {
av_audio_fifo_write(s->fifo, (void **)s->in->extended_data,
s->overlap_skip);
}
s->nb_channels = inlink->channels;
s->chan = av_calloc(inlink->channels, sizeof(*s->chan));
if (!s->chan)
return AVERROR(ENOMEM);
for (i = 0; i < inlink->channels; i++) {
DeclickChannel *c = &s->chan[i];
c->detection = av_calloc(s->window_size, sizeof(*c->detection));
c->auxiliary = av_calloc(s->ar_order + 1, sizeof(*c->auxiliary));
c->acoefficients = av_calloc(s->ar_order + 1, sizeof(*c->acoefficients));
c->acorrelation = av_calloc(s->ar_order + 1, sizeof(*c->acorrelation));
c->tmp = av_calloc(s->ar_order, sizeof(*c->tmp));
c->click = av_calloc(s->window_size, sizeof(*c->click));
c->index = av_calloc(s->window_size, sizeof(*c->index));
c->interpolated = av_calloc(s->window_size, sizeof(*c->interpolated));
if (!c->auxiliary || !c->acoefficients || !c->detection || !c->click ||
!c->index || !c->interpolated || !c->acorrelation || !c->tmp)
return AVERROR(ENOMEM);
}
return 0;
}
static void autocorrelation(const double *input, int order, int size,
double *output, double scale)
{
int i, j;
for (i = 0; i <= order; i++) {
double value = 0.;
for (j = i; j < size; j++)
value += input[j] * input[j - i];
output[i] = value * scale;
}
}
static double autoregression(const double *samples, int ar_order,
int nb_samples, double *k, double *r, double *a)
{
double alpha;
int i, j;
memset(a, 0, ar_order * sizeof(*a));
autocorrelation(samples, ar_order, nb_samples, r, 1. / nb_samples);
/* Levinson-Durbin algorithm */
k[0] = a[0] = -r[1] / r[0];
alpha = r[0] * (1. - k[0] * k[0]);
for (i = 1; i < ar_order; i++) {
double epsilon = 0.;
for (j = 0; j < i; j++)
epsilon += a[j] * r[i - j];
epsilon += r[i + 1];
k[i] = -epsilon / alpha;
alpha *= (1. - k[i] * k[i]);
for (j = i - 1; j >= 0; j--)
k[j] = a[j] + k[i] * a[i - j - 1];
for (j = 0; j <= i; j++)
a[j] = k[j];
}
k[0] = 1.;
for (i = 1; i <= ar_order; i++)
k[i] = a[i - 1];
return sqrt(alpha);
}
static int isfinite_array(double *samples, int nb_samples)
{
int i;
for (i = 0; i < nb_samples; i++)
if (!isfinite(samples[i]))
return 0;
return 1;
}
static int find_index(int *index, int value, int size)
{
int i, start, end;
if ((value < index[0]) || (value > index[size - 1]))
return 1;
i = start = 0;
end = size - 1;
while (start <= end) {
i = (end + start) / 2;
if (index[i] == value)
return 0;
if (value < index[i])
end = i - 1;
if (value > index[i])
start = i + 1;
}
return 1;
}
static int factorization(double *matrix, int n)
{
int i, j, k;
for (i = 0; i < n; i++) {
const int in = i * n;
double value;
value = matrix[in + i];
for (j = 0; j < i; j++)
value -= matrix[j * n + j] * matrix[in + j] * matrix[in + j];
if (value == 0.) {
return -1;
}
matrix[in + i] = value;
for (j = i + 1; j < n; j++) {
const int jn = j * n;
double x;
x = matrix[jn + i];
for (k = 0; k < i; k++)
x -= matrix[k * n + k] * matrix[in + k] * matrix[jn + k];
matrix[jn + i] = x / matrix[in + i];
}
}
return 0;
}
static int do_interpolation(DeclickChannel *c, double *matrix,
double *vector, int n, double *out)
{
int i, j, ret;
double *y;
ret = factorization(matrix, n);
if (ret < 0)
return ret;
av_fast_malloc(&c->y, &c->y_size, n * sizeof(*c->y));
y = c->y;
if (!y)
return AVERROR(ENOMEM);
for (i = 0; i < n; i++) {
const int in = i * n;
double value;
value = vector[i];
for (j = 0; j < i; j++)
value -= matrix[in + j] * y[j];
y[i] = value;
}
for (i = n - 1; i >= 0; i--) {
out[i] = y[i] / matrix[i * n + i];
for (j = i + 1; j < n; j++)
out[i] -= matrix[j * n + i] * out[j];
}
return 0;
}
static int interpolation(DeclickChannel *c, const double *src, int ar_order,
double *acoefficients, int *index, int nb_errors,
double *auxiliary, double *interpolated)
{
double *vector, *matrix;
int i, j;
av_fast_malloc(&c->matrix, &c->matrix_size, nb_errors * nb_errors * sizeof(*c->matrix));
matrix = c->matrix;
if (!matrix)
return AVERROR(ENOMEM);
av_fast_malloc(&c->vector, &c->vector_size, nb_errors * sizeof(*c->vector));
vector = c->vector;
if (!vector)
return AVERROR(ENOMEM);
autocorrelation(acoefficients, ar_order, ar_order + 1, auxiliary, 1.);
for (i = 0; i < nb_errors; i++) {
const int im = i * nb_errors;
for (j = i; j < nb_errors; j++) {
if (abs(index[j] - index[i]) <= ar_order) {
matrix[j * nb_errors + i] = matrix[im + j] = auxiliary[abs(index[j] - index[i])];
} else {
matrix[j * nb_errors + i] = matrix[im + j] = 0;
}
}
}
for (i = 0; i < nb_errors; i++) {
double value = 0.;
for (j = -ar_order; j <= ar_order; j++)
if (find_index(index, index[i] - j, nb_errors))
value -= src[index[i] - j] * auxiliary[abs(j)];
vector[i] = value;
}
return do_interpolation(c, matrix, vector, nb_errors, interpolated);
}
static int detect_clips(AudioDeclickContext *s, DeclickChannel *c,
double unused0,
double *unused1, double *unused2,
uint8_t *clip, int *index,
const double *src, double *dst)
{
const double threshold = s->threshold;
double max_amplitude = 0;
unsigned *histogram;
int i, nb_clips = 0;
av_fast_malloc(&c->histogram, &c->histogram_size, s->nb_hbins * sizeof(*c->histogram));
if (!c->histogram)
return AVERROR(ENOMEM);
histogram = c->histogram;
memset(histogram, 0, sizeof(*histogram) * s->nb_hbins);
for (i = 0; i < s->window_size; i++) {
const unsigned index = fmin(fabs(src[i]), 1) * (s->nb_hbins - 1);
histogram[index]++;
dst[i] = src[i];
clip[i] = 0;
}
for (i = s->nb_hbins - 1; i > 1; i--) {
if (histogram[i]) {
if (histogram[i] / (double)FFMAX(histogram[i - 1], 1) > threshold) {
max_amplitude = i / (double)s->nb_hbins;
}
break;
}
}
if (max_amplitude > 0.) {
for (i = 0; i < s->window_size; i++) {
clip[i] = fabs(src[i]) >= max_amplitude;
}
}
memset(clip, 0, s->ar_order * sizeof(*clip));
memset(clip + (s->window_size - s->ar_order), 0, s->ar_order * sizeof(*clip));
for (i = s->ar_order; i < s->window_size - s->ar_order; i++)
if (clip[i])
index[nb_clips++] = i;
return nb_clips;
}
static int detect_clicks(AudioDeclickContext *s, DeclickChannel *c,
double sigmae,
double *detection, double *acoefficients,
uint8_t *click, int *index,
const double *src, double *dst)
{
const double threshold = s->threshold;
int i, j, nb_clicks = 0, prev = -1;
memset(detection, 0, s->window_size * sizeof(*detection));
for (i = s->ar_order; i < s->window_size; i++) {
for (j = 0; j <= s->ar_order; j++) {
detection[i] += acoefficients[j] * src[i - j];
}
}
for (i = 0; i < s->window_size; i++) {
click[i] = fabs(detection[i]) > sigmae * threshold;
dst[i] = src[i];
}
for (i = 0; i < s->window_size; i++) {
if (!click[i])
continue;
if (prev >= 0 && (i > prev + 1) && (i <= s->nb_burst_samples + prev))
for (j = prev + 1; j < i; j++)
click[j] = 1;
prev = i;
}
memset(click, 0, s->ar_order * sizeof(*click));
memset(click + (s->window_size - s->ar_order), 0, s->ar_order * sizeof(*click));
for (i = s->ar_order; i < s->window_size - s->ar_order; i++)
if (click[i])
index[nb_clicks++] = i;
return nb_clicks;
}
typedef struct ThreadData {
AVFrame *out;
} ThreadData;
static int filter_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
{
AudioDeclickContext *s = ctx->priv;
ThreadData *td = arg;
AVFrame *out = td->out;
const double *src = (const double *)s->in->extended_data[ch];
double *is = (double *)s->is->extended_data[ch];
double *dst = (double *)s->out->extended_data[ch];
double *ptr = (double *)out->extended_data[ch];
double *buf = (double *)s->buffer->extended_data[ch];
const double *w = s->window_func_lut;
DeclickChannel *c = &s->chan[ch];
double sigmae;
int j, ret;
sigmae = autoregression(src, s->ar_order, s->window_size, c->acoefficients, c->acorrelation, c->tmp);
if (isfinite_array(c->acoefficients, s->ar_order + 1)) {
double *interpolated = c->interpolated;
int *index = c->index;
int nb_errors;
nb_errors = s->detector(s, c, sigmae, c->detection, c->acoefficients,
c->click, index, src, dst);
if (nb_errors > 0) {
ret = interpolation(c, src, s->ar_order, c->acoefficients, index,
nb_errors, c->auxiliary, interpolated);
if (ret < 0)
return ret;
for (j = 0; j < nb_errors; j++) {
dst[index[j]] = interpolated[j];
is[index[j]] = 1;
}
}
} else {
memcpy(dst, src, s->window_size * sizeof(*dst));
}
if (s->method == 0) {
for (j = 0; j < s->window_size; j++)
buf[j] += dst[j] * w[j];
} else {
const int skip = s->overlap_skip;
for (j = 0; j < s->hop_size; j++)
buf[j] = dst[skip + j];
}
for (j = 0; j < s->hop_size; j++)
ptr[j] = buf[j];
memmove(buf, buf + s->hop_size, (s->window_size * 2 - s->hop_size) * sizeof(*buf));
memmove(is, is + s->hop_size, (s->window_size - s->hop_size) * sizeof(*is));
memset(buf + s->window_size * 2 - s->hop_size, 0, s->hop_size * sizeof(*buf));
memset(is + s->window_size - s->hop_size, 0, s->hop_size * sizeof(*is));
return 0;
}
static int filter_frame(AVFilterLink *inlink, AVFrame *in)
{
AVFilterContext *ctx = inlink->dst;
AVFilterLink *outlink = ctx->outputs[0];
AudioDeclickContext *s = ctx->priv;
AVFrame *out = NULL;
int ret = 0;
if (s->pts == AV_NOPTS_VALUE)
s->pts = in->pts;
ret = av_audio_fifo_write(s->fifo, (void **)in->extended_data,
in->nb_samples);
av_frame_free(&in);
while (av_audio_fifo_size(s->fifo) >= s->window_size) {
int j, ch, detected_errors = 0;
ThreadData td;
out = ff_get_audio_buffer(outlink, s->hop_size);
if (!out)
return AVERROR(ENOMEM);
ret = av_audio_fifo_peek(s->fifo, (void **)s->in->extended_data,
s->window_size);
if (ret < 0)
break;
td.out = out;
ret = ctx->internal->execute(ctx, filter_channel, &td, NULL, inlink->channels);
if (ret < 0)
goto fail;
for (ch = 0; ch < s->in->channels; ch++) {
double *is = (double *)s->is->extended_data[ch];
for (j = 0; j < s->hop_size; j++) {
if (is[j])
detected_errors++;
}
}
av_audio_fifo_drain(s->fifo, s->hop_size);
if (s->samples_left > 0)
out->nb_samples = FFMIN(s->hop_size, s->samples_left);
out->pts = s->pts;
s->pts += s->hop_size;
s->detected_errors += detected_errors;
s->nb_samples += out->nb_samples * inlink->channels;
ret = ff_filter_frame(outlink, out);
if (ret < 0)
break;
if (s->samples_left > 0) {
s->samples_left -= s->hop_size;
if (s->samples_left <= 0)
av_audio_fifo_drain(s->fifo, av_audio_fifo_size(s->fifo));
}
}
fail:
if (ret < 0)
av_frame_free(&out);
return ret;
}
static int request_frame(AVFilterLink *outlink)
{
AVFilterContext *ctx = outlink->src;
AudioDeclickContext *s = ctx->priv;
int ret = 0;
ret = ff_request_frame(ctx->inputs[0]);
if (ret == AVERROR_EOF && av_audio_fifo_size(s->fifo) > 0) {
if (!s->samples_left)
s->samples_left = av_audio_fifo_size(s->fifo) - s->overlap_skip;
if (s->samples_left > 0) {
AVFrame *in = ff_get_audio_buffer(outlink, s->window_size - s->samples_left);
if (!in)
return AVERROR(ENOMEM);
ret = filter_frame(ctx->inputs[0], in);
}
}
return ret;
}
static av_cold int init(AVFilterContext *ctx)
{
AudioDeclickContext *s = ctx->priv;
s->is_declip = !strcmp(ctx->filter->name, "adeclip");
if (s->is_declip) {
s->detector = detect_clips;
} else {
s->detector = detect_clicks;
}
return 0;
}
static av_cold void uninit(AVFilterContext *ctx)
{
AudioDeclickContext *s = ctx->priv;
int i;
av_log(ctx, AV_LOG_INFO, "Detected %s in %"PRId64" of %"PRId64" samples (%g%%).\n",
s->is_declip ? "clips" : "clicks", s->detected_errors,
s->nb_samples, 100. * s->detected_errors / s->nb_samples);
av_audio_fifo_free(s->fifo);
av_freep(&s->window_func_lut);
av_frame_free(&s->in);
av_frame_free(&s->out);
av_frame_free(&s->buffer);
av_frame_free(&s->is);
if (s->chan) {
for (i = 0; i < s->nb_channels; i++) {
DeclickChannel *c = &s->chan[i];
av_freep(&c->detection);
av_freep(&c->auxiliary);
av_freep(&c->acoefficients);
av_freep(&c->acorrelation);
av_freep(&c->tmp);
av_freep(&c->click);
av_freep(&c->index);
av_freep(&c->interpolated);
av_freep(&c->matrix);
c->matrix_size = 0;
av_freep(&c->histogram);
c->histogram_size = 0;
av_freep(&c->vector);
c->vector_size = 0;
av_freep(&c->y);
c->y_size = 0;
}
}
av_freep(&s->chan);
s->nb_channels = 0;
}
static const AVFilterPad inputs[] = {
{
.name = "default",
.type = AVMEDIA_TYPE_AUDIO,
.filter_frame = filter_frame,
.config_props = config_input,
},
{ NULL }
};
static const AVFilterPad outputs[] = {
{
.name = "default",
.type = AVMEDIA_TYPE_AUDIO,
.request_frame = request_frame,
},
{ NULL }
};
AVFilter ff_af_adeclick = {
.name = "adeclick",
.description = NULL_IF_CONFIG_SMALL("Remove impulsive noise from input audio."),
.query_formats = query_formats,
.priv_size = sizeof(AudioDeclickContext),
.priv_class = &adeclick_class,
.init = init,
.uninit = uninit,
.inputs = inputs,
.outputs = outputs,
.flags = AVFILTER_FLAG_SLICE_THREADS,
};
static const AVOption adeclip_options[] = {
{ "w", "set window size", OFFSET(w), AV_OPT_TYPE_DOUBLE, {.dbl=55}, 10, 100, AF },
{ "o", "set window overlap", OFFSET(overlap), AV_OPT_TYPE_DOUBLE, {.dbl=75}, 50, 95, AF },
{ "a", "set autoregression order", OFFSET(ar), AV_OPT_TYPE_DOUBLE, {.dbl=8}, 0, 25, AF },
{ "t", "set threshold", OFFSET(threshold), AV_OPT_TYPE_DOUBLE, {.dbl=10}, 1, 100, AF },
{ "n", "set histogram size", OFFSET(nb_hbins), AV_OPT_TYPE_INT, {.i64=1000}, 100, 9999, AF },
{ "m", "set overlap method", OFFSET(method), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, AF, "m" },
{ "a", "overlap-add", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, AF, "m" },
{ "s", "overlap-save", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, AF, "m" },
{ NULL }
};
AVFILTER_DEFINE_CLASS(adeclip);
AVFilter ff_af_adeclip = {
.name = "adeclip",
.description = NULL_IF_CONFIG_SMALL("Remove clipping from input audio."),
.query_formats = query_formats,
.priv_size = sizeof(AudioDeclickContext),
.priv_class = &adeclip_class,
.init = init,
.uninit = uninit,
.inputs = inputs,
.outputs = outputs,
.flags = AVFILTER_FLAG_SLICE_THREADS,
}; |