libavfilter/vf_sr.c

   1 /*
   2  * Copyright (c) 2018 Sergey Lavrushkin
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 /**
  22  * @file
  23  * Filter implementing image super-resolution using deep convolutional networks.
  24  * https://arxiv.org/abs/1501.00092
  25  * https://arxiv.org/abs/1609.05158
  26  */
  27
  28 #include "avfilter.h"
  29 #include "filters.h"
  30 #include "video.h"
  31 #include "libavutil/opt.h"
  32 #include "libavutil/pixdesc.h"
  33 #include "libswscale/swscale.h"
  34 #include "dnn_filter_common.h"
  35
  36 typedef struct SRContext {
  37     const AVClass *class;
  38     DnnContext dnnctx;
  39     int scale_factor;
  40     struct SwsContext *sws_uv_scale;
  41     int sws_uv_height;
  42     struct SwsContext *sws_pre_scale;
  43 } SRContext;
  44
  45 #define OFFSET(x) offsetof(SRContext, x)
  46 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
  47 static const AVOption sr_options[] = {
  48     { "dnn_backend", "DNN backend used for model execution", OFFSET(dnnctx.backend_type), AV_OPT_TYPE_INT, { .i64 = 1 }, 0, 1, FLAGS, .unit = "backend" },
  49 #if (CONFIG_LIBTENSORFLOW == 1)
  50     { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, .unit = "backend" },
  51 #endif
  52     { "scale_factor", "scale factor for SRCNN model", OFFSET(scale_factor), AV_OPT_TYPE_INT, { .i64 = 2 }, 2, 4, FLAGS },
  53     { NULL }
  54 };
  55
  56 AVFILTER_DNN_DEFINE_CLASS(sr, DNN_TF);
  57
  58 static av_cold int init(AVFilterContext *context)
  59 {
  60     SRContext *sr_context = context->priv;
  61     return ff_dnn_init(&sr_context->dnnctx, DFT_PROCESS_FRAME, context);
  62 }
  63
  64 static const enum AVPixelFormat pixel_formats[] = {
  65     AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
  66     AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_GRAY8,
  67     AV_PIX_FMT_NONE
  68 };
  69
  70 static int config_output(AVFilterLink *outlink)
  71 {
  72     AVFilterContext *context = outlink->src;
  73     SRContext *ctx = context->priv;
  74     int result;
  75     AVFilterLink *inlink = context->inputs[0];
  76     int out_width, out_height;
  77
  78     // have a try run in case that the dnn model resize the frame
  79     result = ff_dnn_get_output(&ctx->dnnctx, inlink->w, inlink->h, &out_width, &out_height);
  80     if (result != 0) {
  81         av_log(ctx, AV_LOG_ERROR, "could not get output from the model\n");
  82         return result;
  83     }
  84
  85     if (inlink->w != out_width || inlink->h != out_height) {
  86         //espcn
  87         outlink->w = out_width;
  88         outlink->h = out_height;
  89         if (inlink->format != AV_PIX_FMT_GRAY8){
  90             const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
  91             int sws_src_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
  92             int sws_src_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
  93             int sws_dst_h = AV_CEIL_RSHIFT(outlink->h, desc->log2_chroma_h);
  94             int sws_dst_w = AV_CEIL_RSHIFT(outlink->w, desc->log2_chroma_w);
  95             ctx->sws_uv_scale = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
  96                                                sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
  97                                                SWS_BICUBIC, NULL, NULL, NULL);
  98             ctx->sws_uv_height = sws_src_h;
  99         }
 100     } else {
 101         //srcnn
 102         outlink->w = out_width * ctx->scale_factor;
 103         outlink->h = out_height * ctx->scale_factor;
 104         ctx->sws_pre_scale = sws_getContext(inlink->w, inlink->h, inlink->format,
 105                                         outlink->w, outlink->h, outlink->format,
 106                                         SWS_BICUBIC, NULL, NULL, NULL);
 107     }
 108
 109     return 0;
 110 }
 111
 112 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
 113 {
 114     DNNAsyncStatusType async_state = 0;
 115     AVFilterContext *context = inlink->dst;
 116     SRContext *ctx = context->priv;
 117     AVFilterLink *outlink = context->outputs[0];
 118     AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 119     int dnn_result;
 120
 121     if (!out){
 122         av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
 123         av_frame_free(&in);
 124         return AVERROR(ENOMEM);
 125     }
 126     av_frame_copy_props(out, in);
 127
 128     if (ctx->sws_pre_scale) {
 129         sws_scale(ctx->sws_pre_scale,
 130                     (const uint8_t **)in->data, in->linesize, 0, in->height,
 131                     out->data, out->linesize);
 132         dnn_result = ff_dnn_execute_model(&ctx->dnnctx, out, out);
 133     } else {
 134         dnn_result = ff_dnn_execute_model(&ctx->dnnctx, in, out);
 135     }
 136
 137     if (dnn_result != 0){
 138         av_log(ctx, AV_LOG_ERROR, "failed to execute loaded model\n");
 139         av_frame_free(&in);
 140         av_frame_free(&out);
 141         return dnn_result;
 142     }
 143
 144     do {
 145         async_state = ff_dnn_get_result(&ctx->dnnctx, &in, &out);
 146     } while (async_state == DAST_NOT_READY);
 147
 148     if (async_state != DAST_SUCCESS)
 149         return AVERROR(EINVAL);
 150
 151     if (ctx->sws_uv_scale) {
 152         sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 1), in->linesize + 1,
 153                   0, ctx->sws_uv_height, out->data + 1, out->linesize + 1);
 154         sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 2), in->linesize + 2,
 155                   0, ctx->sws_uv_height, out->data + 2, out->linesize + 2);
 156     }
 157     if (in != out) {
 158         av_frame_free(&in);
 159     }
 160     return ff_filter_frame(outlink, out);
 161 }
 162
 163 static av_cold void uninit(AVFilterContext *context)
 164 {
 165     SRContext *sr_context = context->priv;
 166
 167     ff_dnn_uninit(&sr_context->dnnctx);
 168     sws_freeContext(sr_context->sws_uv_scale);
 169     sws_freeContext(sr_context->sws_pre_scale);
 170 }
 171
 172 static const AVFilterPad sr_inputs[] = {
 173     {
 174         .name         = "default",
 175         .type         = AVMEDIA_TYPE_VIDEO,
 176         .filter_frame = filter_frame,
 177     },
 178 };
 179
 180 static const AVFilterPad sr_outputs[] = {
 181     {
 182         .name = "default",
 183         .config_props = config_output,
 184         .type = AVMEDIA_TYPE_VIDEO,
 185     },
 186 };
 187
 188 const FFFilter ff_vf_sr = {
 189     .p.name        = "sr",
 190     .p.description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
 191     .p.priv_class  = &sr_class,
 192     .priv_size     = sizeof(SRContext),
 193     .preinit       = ff_dnn_filter_init_child_class,
 194     .init          = init,
 195     .uninit        = uninit,
 196     FILTER_INPUTS(sr_inputs),
 197     FILTER_OUTPUTS(sr_outputs),
 198     FILTER_PIXFMTS_ARRAY(pixel_formats),
 199 };