视频降噪算法 hqdn3d 原理分析

DogDaoDao2024-05-13 10:28

视频降噪

视频降噪是一种处理技术，旨在减少视频中的噪声，提高画面质量。噪声可能来自多种源头，包括摄像机的传感器、压缩算法、传输过程中的干扰等。降噪处理对于视频监控、视频会议、电影后期制作以及任何需要高画质输出的应用场景都非常重要。

视频降噪技术通常包括以下几个关键点：

  • 时域降噪：利用连续视频帧之间的时间冗余，通过分析连续帧中相同位置像素的变化来识别噪声，并进行相应的处理。

  • 空域降噪：在单帧视频内，通过分析像素邻域内的亮度和色度信息来识别并减少噪声。

  • 3D降噪：结合时域和空域信息，同时考虑视频帧的当前帧和相邻帧，进行更加复杂的降噪处理。

  • 非线性降噪：某些算法采用非线性处理方法，如中值滤波或维纳滤波，来减少噪声，这些方法对某些类型的噪声特别有效。

  • 自适应降噪：根据视频内容的复杂性和噪声水平动态调整降噪强度。

  • 基于模型的降噪：使用统计模型或机器学习技术来识别噪声模式，并进行相应的降噪处理。

  • 运动补偿：在视频中，运动物体的噪声处理需要特别考虑，以避免在运动补偿过程中引入伪影。

  • 硬件加速：某些降噪算法可以在GPU或其他专用硬件上执行，以提高处理速度。

  • 用户控制：许多降噪工具提供用户界面，允许用户根据需要调整降噪参数。

关于视频降噪的算法综述可以参考：视频降噪综述。

hqdn3d

hqdn3d（High Quality 3D Denoise）是一种高质量的三维视频降噪算法，主要用于去除视频中的噪声。hqdn3d算法能够有效地去除视频中的高频率噪声，同时尽量减少图像的失真和模糊，保持视频质量。

hqdn3d算法已经被实现为AviSynth插件，并且可以在FFmpeg中作为滤镜使用，例如通过命令 ffmpeg -i input.avi -vf hqdn3d output.mp4 来应用。高质量去噪3D是一个AviSynth端口的同名MPlayer过滤器。采用3路低通滤波，在完全去除高频噪声的同时最大限度地减少混合伪影。

原理

hqdn3d 算法的原理相对比较容易理解，总结其原理，如下：

  • 该算法是像素处理，不涉及频域和运动搜索；
  • 空域去噪时，当前像素参考当前帧左邻、上邻和前一帧同一位置的像素，对每个像素多次应用低通滤波器；
  • 时域去噪时，参考前一帧同一位置的像素，对每个像素应用一个低通滤波器；
  • 根据相邻像素的差，确定偏移量，将相邻像素颜色值"拉近"，即对像素做加减法，以达到降噪目的；
  • 对于较小的像素差异，hqdn3d会拉近距离，若差异较大，降低偏置系数，达到保边效果；
  • 可以通过外部设置滤波强度，调整峰值位置；
  1. 分析 FFmpeg 中 hqdn3d 算法，算法实现流程如下：

源码

  • FFmpeg 中 filter 滤镜对应的 hqdn3d 算法实现
c 复制代码 
/*
 * Copyright (c) 2003 Daniel Moreno <comac AT comac DOT darktech DOT org>
 * Copyright (c) 2010 Baptiste Coudurier
 * Copyright (c) 2012 Loren Merritt
 *
 * This file is part of FFmpeg, ported from MPlayer.
 *
 * FFmpeg is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 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 General Public License for more details.
 *
 * You should have received a copy of the GNU 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
 * high quality 3d video denoiser, ported from MPlayer
 * libmpcodecs/vf_hqdn3d.c.
 */

#include <float.h>

#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/common.h"
#include "libavutil/pixdesc.h"
#include "libavutil/intreadwrite.h"
#include "libavutil/opt.h"

#include "avfilter.h"
#include "formats.h"
#include "internal.h"
#include "video.h"
#include "vf_hqdn3d.h"

#define LUT_BITS (depth==16 ? 8 : 4)
#define LOAD(x) (((depth == 8 ? src[x] : AV_RN16A(src + (x) * 2)) << (16 - depth))\
                 + (((1 << (16 - depth)) - 1) >> 1))
#define STORE(x,val) (depth == 8 ? dst[x] = (val) >> (16 - depth) : \
                                   AV_WN16A(dst + (x) * 2, (val) >> (16 - depth)))

av_always_inline
static uint32_t lowpass(int prev, int cur, int16_t *coef, int depth)
{
    int d = (prev - cur) >> (8 - LUT_BITS);
    return cur + coef[d];
}

av_always_inline
static void denoise_temporal(uint8_t *src, uint8_t *dst,
                             uint16_t *frame_ant,
                             int w, int h, int sstride, int dstride,
                             int16_t *temporal, int depth)
{
    long x, y;
    uint32_t tmp;

    temporal += 256 << LUT_BITS;

    for (y = 0; y < h; y++) {
        for (x = 0; x < w; x++) {
            frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal, depth);
            STORE(x, tmp);
        }
        src += sstride;
        dst += dstride;
        frame_ant += w;
    }
}

av_always_inline
static void denoise_spatial(HQDN3DContext *s,
                            uint8_t *src, uint8_t *dst,
                            uint16_t *line_ant, uint16_t *frame_ant,
                            int w, int h, int sstride, int dstride,
                            int16_t *spatial, int16_t *temporal, int depth)
{
    long x, y;
    uint32_t pixel_ant;
    uint32_t tmp;

    spatial  += 256 << LUT_BITS;
    temporal += 256 << LUT_BITS;

    /* First line has no top neighbor. Only left one for each tmp and
     * last frame */
    pixel_ant = LOAD(0);
    for (x = 0; x < w; x++) {
        line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial, depth);
        frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
        STORE(x, tmp);
    }

    for (y = 1; y < h; y++) {
        src += sstride;
        dst += dstride;
        frame_ant += w;
        if (s->denoise_row[depth]) {
            s->denoise_row[depth](src, dst, line_ant, frame_ant, w, spatial, temporal);
            continue;
        }
        pixel_ant = LOAD(0);
        for (x = 0; x < w-1; x++) {
            line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
            pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial, depth);
            frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
            STORE(x, tmp);
        }
        line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
        frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
        STORE(x, tmp);
    }
}

av_always_inline
static int denoise_depth(HQDN3DContext *s,
                         uint8_t *src, uint8_t *dst,
                         uint16_t *line_ant, uint16_t **frame_ant_ptr,
                         int w, int h, int sstride, int dstride,
                         int16_t *spatial, int16_t *temporal, int depth)
{
    // FIXME: For 16-bit depth, frame_ant could be a pointer to the previous
    // filtered frame rather than a separate buffer.
    long x, y;
    uint16_t *frame_ant = *frame_ant_ptr;
    if (!frame_ant) {
        uint8_t *frame_src = src;
        *frame_ant_ptr = frame_ant = av_malloc_array(w, h*sizeof(uint16_t));
        if (!frame_ant)
            return AVERROR(ENOMEM);
        for (y = 0; y < h; y++, src += sstride, frame_ant += w)
            for (x = 0; x < w; x++)
                frame_ant[x] = LOAD(x);
        src = frame_src;
        frame_ant = *frame_ant_ptr;
    }

    if (spatial[0])
        denoise_spatial(s, src, dst, line_ant, frame_ant,
                        w, h, sstride, dstride, spatial, temporal, depth);
    else
        denoise_temporal(src, dst, frame_ant,
                         w, h, sstride, dstride, temporal, depth);
    emms_c();
    return 0;
}

#define denoise(...)                                                          \
    do {                                                                      \
        int ret = AVERROR_BUG;                                                \
        switch (s->depth) {                                                   \
            case  8: ret = denoise_depth(__VA_ARGS__,  8); break;             \
            case  9: ret = denoise_depth(__VA_ARGS__,  9); break;             \
            case 10: ret = denoise_depth(__VA_ARGS__, 10); break;             \
            case 12: ret = denoise_depth(__VA_ARGS__, 12); break;             \
            case 14: ret = denoise_depth(__VA_ARGS__, 14); break;             \
            case 16: ret = denoise_depth(__VA_ARGS__, 16); break;             \
        }                                                                     \
        if (ret < 0) {                                                        \
            av_frame_free(&out);                                              \
            if (!direct)                                                      \
                av_frame_free(&in);                                           \
            return ret;                                                       \
        }                                                                     \
    } while (0)

static void precalc_coefs(double dist25, int depth, int16_t *ct)
{
    int i;
    double gamma, simil, C;

    gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.00001);

    for (i = -256<<LUT_BITS; i < 256<<LUT_BITS; i++) {
        double f = ((i<<(9-LUT_BITS)) + (1<<(8-LUT_BITS)) - 1) / 512.0; // midpoint of the bin
        simil = FFMAX(0, 1.0 - fabs(f) / 255.0);
        C = pow(simil, gamma) * 256.0 * f;
        ct[(256<<LUT_BITS)+i] = lrint(C);
    }

    ct[0] = !!dist25;
}

#define PARAM1_DEFAULT 4.0
#define PARAM2_DEFAULT 3.0
#define PARAM3_DEFAULT 6.0

static av_cold int init(AVFilterContext *ctx)
{
    HQDN3DContext *s = ctx->priv;

    if (!s->strength[LUMA_SPATIAL])
        s->strength[LUMA_SPATIAL] = PARAM1_DEFAULT;
    if (!s->strength[CHROMA_SPATIAL])
        s->strength[CHROMA_SPATIAL] = PARAM2_DEFAULT * s->strength[LUMA_SPATIAL] / PARAM1_DEFAULT;
    if (!s->strength[LUMA_TMP])
        s->strength[LUMA_TMP]   = PARAM3_DEFAULT * s->strength[LUMA_SPATIAL] / PARAM1_DEFAULT;
    if (!s->strength[CHROMA_TMP])
        s->strength[CHROMA_TMP] = s->strength[LUMA_TMP] * s->strength[CHROMA_SPATIAL] / s->strength[LUMA_SPATIAL];

    av_log(ctx, AV_LOG_VERBOSE, "ls:%f cs:%f lt:%f ct:%f\n",
           s->strength[LUMA_SPATIAL], s->strength[CHROMA_SPATIAL],
           s->strength[LUMA_TMP], s->strength[CHROMA_TMP]);

    return 0;
}

static av_cold void uninit(AVFilterContext *ctx)
{
    HQDN3DContext *s = ctx->priv;

    av_freep(&s->coefs[0]);
    av_freep(&s->coefs[1]);
    av_freep(&s->coefs[2]);
    av_freep(&s->coefs[3]);
    av_freep(&s->line[0]);
    av_freep(&s->line[1]);
    av_freep(&s->line[2]);
    av_freep(&s->frame_prev[0]);
    av_freep(&s->frame_prev[1]);
    av_freep(&s->frame_prev[2]);
}

static int query_formats(AVFilterContext *ctx)
{
    static const enum AVPixelFormat pix_fmts[] = {
        AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
        AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV440P,
        AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
        AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
        AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
        AV_PIX_FMT_YUV440P10,
        AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
        AV_PIX_FMT_YUV440P12,
        AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
        AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
        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);
}

static void calc_coefs(AVFilterContext *ctx)
{
    HQDN3DContext *s = ctx->priv;

    for (int i = 0; i < 4; i++)
        precalc_coefs(s->strength[i], s->depth, s->coefs[i]);
}

static int config_input(AVFilterLink *inlink)
{
    AVFilterContext *ctx = inlink->dst;
    HQDN3DContext *s = inlink->dst->priv;
    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
    int i, depth;

    uninit(inlink->dst);

    s->hsub  = desc->log2_chroma_w;
    s->vsub  = desc->log2_chroma_h;
    s->depth = depth = desc->comp[0].depth;

    for (i = 0; i < 3; i++) {
        s->line[i] = av_malloc_array(inlink->w, sizeof(*s->line[i]));
        if (!s->line[i])
            return AVERROR(ENOMEM);
    }

    for (i = 0; i < 4; i++) {
        s->coefs[i] = av_malloc((512<<LUT_BITS) * sizeof(int16_t));
        if (!s->coefs[i])
            return AVERROR(ENOMEM);
    }

    calc_coefs(ctx);

    if (ARCH_X86)
        ff_hqdn3d_init_x86(s);

    return 0;
}

typedef struct ThreadData {
    AVFrame *in, *out;
    int direct;
} ThreadData;

static int do_denoise(AVFilterContext *ctx, void *data, int job_nr, int n_jobs)
{
    HQDN3DContext *s = ctx->priv;
    const ThreadData *td = data;
    AVFrame *out = td->out;
    AVFrame *in = td->in;
    int direct = td->direct;

    denoise(s, in->data[job_nr], out->data[job_nr],
                s->line[job_nr], &s->frame_prev[job_nr],
                AV_CEIL_RSHIFT(in->width,  (!!job_nr * s->hsub)),
                AV_CEIL_RSHIFT(in->height, (!!job_nr * s->vsub)),
                in->linesize[job_nr], out->linesize[job_nr],
                s->coefs[job_nr ? CHROMA_SPATIAL : LUMA_SPATIAL],
                s->coefs[job_nr ? CHROMA_TMP     : LUMA_TMP]);

    return 0;
}

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

    AVFrame *out;
    int direct = av_frame_is_writable(in) && !ctx->is_disabled;
    ThreadData td;

    if (direct) {
        out = in;
    } else {
        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);
    }

    td.in = in;
    td.out = out;
    td.direct = direct;
    /* one thread per plane */
    ctx->internal->execute(ctx, do_denoise, &td, NULL, 3);

    if (ctx->is_disabled) {
        av_frame_free(&out);
        return ff_filter_frame(outlink, in);
    }

    if (!direct)
        av_frame_free(&in);

    return ff_filter_frame(outlink, out);
}

static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
                           char *res, int res_len, int flags)
{
    int ret;

    ret = ff_filter_process_command(ctx, cmd, args, res, res_len, flags);
    if (ret < 0)
        return ret;

    calc_coefs(ctx);

    return 0;
}

#define OFFSET(x) offsetof(HQDN3DContext, x)
#define FLAGS AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_RUNTIME_PARAM
static const AVOption hqdn3d_options[] = {
    { "luma_spatial",   "spatial luma strength",    OFFSET(strength[LUMA_SPATIAL]),   AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
    { "chroma_spatial", "spatial chroma strength",  OFFSET(strength[CHROMA_SPATIAL]), AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
    { "luma_tmp",       "temporal luma strength",   OFFSET(strength[LUMA_TMP]),       AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
    { "chroma_tmp",     "temporal chroma strength", OFFSET(strength[CHROMA_TMP]),     AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
    { NULL }
};

AVFILTER_DEFINE_CLASS(hqdn3d);

static const AVFilterPad avfilter_vf_hqdn3d_inputs[] = {
    {
        .name         = "default",
        .type         = AVMEDIA_TYPE_VIDEO,
        .config_props = config_input,
        .filter_frame = filter_frame,
    },
    { NULL }
};


static const AVFilterPad avfilter_vf_hqdn3d_outputs[] = {
    {
        .name = "default",
        .type = AVMEDIA_TYPE_VIDEO
    },
    { NULL }
};

AVFilter ff_vf_hqdn3d = {
    .name          = "hqdn3d",
    .description   = NULL_IF_CONFIG_SMALL("Apply a High Quality 3D Denoiser."),
    .priv_size     = sizeof(HQDN3DContext),
    .priv_class    = &hqdn3d_class,
    .init          = init,
    .uninit        = uninit,
    .query_formats = query_formats,
    .inputs        = avfilter_vf_hqdn3d_inputs,
    .outputs       = avfilter_vf_hqdn3d_outputs,
    .flags         = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
    .process_command = process_command,
};
相关推荐
安静读书17 分钟前
Python解析视频FPS(帧率)、分辨率信息
python·opencv·音视频
佑华硬盘拷贝机19 分钟前
音频档案批量拷贝：专业SD拷贝机解决方案
音视频
EasyNVR24 分钟前
NVR管理平台EasyNVR多个NVR同时管理：全方位安防监控视频融合云平台方案
安全·音视频·监控·视频监控
xcLeigh8 小时前
HTML5超酷响应式视频背景动画特效(六种风格，附源码)
前端·音视频·html5
韩曙亮9 小时前
【FFmpeg】FFmpeg 内存结构 ③ ( AVPacket 函数简介 | av_packet_ref 函数 | av_packet_clone 函数 )
ffmpeg·音视频·avpacket·av_packet_clone·av_packet_ref·ffmpeg内存结构
9527华安13 小时前
FPGA实现PCIE3.0视频采集转10G万兆UDP网络输出，基于XDMA+GTH架构，提供工程源码和技术支持
网络·fpga开发·udp·音视频·xdma·pcie3.0·万兆网
电子科技圈14 小时前
XMOS携手合作伙伴晓龙国际联合推出集成了ASRC等功能的多通道音频板
科技·嵌入式硬件·mcu·物联网·音视频·iot
oushaojun214 小时前
ubuntu中使用ffmpeg和nginx推流rtmp视频
nginx·ubuntu·ffmpeg·rtmp
码码哈哈0.014 小时前
免费的视频混剪综合处理工具介绍与下载
音视频
莫固执，朋友14 小时前
网络抓包工具tcpdump 在海思平台上的编译使用
网络·ffmpeg·音视频·tcpdump
热门推荐
01(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明02PyTorch机器学习实现液态神经网络03【HarmonyOS】HUAWEI DevEco Studio 下载地址汇总04玄机平台应急响应—webshell查杀05Coze扣子平台完整体验和实践（附国内和国际版对比）06Ubuntu 20.04使用Livox mid 360 测试 FAST_LIO07【目标跟踪】相机运动补偿08Unity中PICO实现 隔空取物 和 接触抓取物体09RAG 实践- Ollama+RagFlow 部署本地知识库10怎样让音频速度变慢？请跟随以下方法进行操作