2 * Copyright (c) 2018 Sergey Lavrushkin
4 * This file is part of FFmpeg.
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.
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.
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
23 * Filter implementing image super-resolution using deep convolutional networks.
24 * https://arxiv.org/abs/1501.00092
25 * https://arxiv.org/abs/1609.05158
31 #include "libavutil/opt.h"
32 #include "libavutil/pixdesc.h"
33 #include "libavformat/avio.h"
34 #include "libswscale/swscale.h"
35 #include "dnn_interface.h"
37 typedef struct SRContext {
41 DNNBackendType backend_type;
42 DNNModule *dnn_module;
45 struct SwsContext *sws_uv_scale;
47 struct SwsContext *sws_pre_scale;
50 #define OFFSET(x) offsetof(SRContext, x)
51 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
52 static const AVOption sr_options[] = {
53 { "dnn_backend", "DNN backend used for model execution", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS, "backend" },
54 { "native", "native backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, 0, 0, FLAGS, "backend" },
55 #if (CONFIG_LIBTENSORFLOW == 1)
56 { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, "backend" },
58 { "scale_factor", "scale factor for SRCNN model", OFFSET(scale_factor), AV_OPT_TYPE_INT, { .i64 = 2 }, 2, 4, FLAGS },
59 { "model", "path to model file specifying network architecture and its parameters", OFFSET(model_filename), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
63 AVFILTER_DEFINE_CLASS(sr);
65 static av_cold int init(AVFilterContext *context)
67 SRContext *sr_context = context->priv;
69 sr_context->dnn_module = ff_get_dnn_module(sr_context->backend_type);
70 if (!sr_context->dnn_module){
71 av_log(context, AV_LOG_ERROR, "could not create DNN module for requested backend\n");
72 return AVERROR(ENOMEM);
75 if (!sr_context->model_filename){
76 av_log(context, AV_LOG_ERROR, "model file for network was not specified\n");
79 if (!sr_context->dnn_module->load_model) {
80 av_log(context, AV_LOG_ERROR, "load_model for network was not specified\n");
83 sr_context->model = (sr_context->dnn_module->load_model)(sr_context->model_filename, NULL, NULL);
84 if (!sr_context->model){
85 av_log(context, AV_LOG_ERROR, "could not load DNN model\n");
92 static int query_formats(AVFilterContext *context)
94 const enum AVPixelFormat pixel_formats[] = {AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
95 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_GRAY8,
97 AVFilterFormats *formats_list;
99 formats_list = ff_make_format_list(pixel_formats);
101 av_log(context, AV_LOG_ERROR, "could not create formats list\n");
102 return AVERROR(ENOMEM);
105 return ff_set_common_formats(context, formats_list);
108 static int config_output(AVFilterLink *outlink)
110 AVFilterContext *context = outlink->src;
111 SRContext *ctx = context->priv;
112 DNNReturnType result;
113 AVFilterLink *inlink = context->inputs[0];
115 const char *model_output_name = "y";
117 // have a try run in case that the dnn model resize the frame
118 AVFrame *fake_in = ff_get_video_buffer(inlink, inlink->w, inlink->h);
119 out = ff_get_video_buffer(inlink, inlink->w, inlink->h);
120 result = (ctx->dnn_module->execute_model)(ctx->model, "x", fake_in,
121 (const char **)&model_output_name, 1, out);
122 if (result != DNN_SUCCESS){
123 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
127 if (fake_in->width != out->width || fake_in->height != out->height) {
129 outlink->w = out->width;
130 outlink->h = out->height;
131 if (inlink->format != AV_PIX_FMT_GRAY8){
132 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
133 int sws_src_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
134 int sws_src_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
135 int sws_dst_h = AV_CEIL_RSHIFT(outlink->h, desc->log2_chroma_h);
136 int sws_dst_w = AV_CEIL_RSHIFT(outlink->w, desc->log2_chroma_w);
137 ctx->sws_uv_scale = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
138 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
139 SWS_BICUBIC, NULL, NULL, NULL);
140 ctx->sws_uv_height = sws_src_h;
144 outlink->w = out->width * ctx->scale_factor;
145 outlink->h = out->height * ctx->scale_factor;
146 ctx->sws_pre_scale = sws_getContext(inlink->w, inlink->h, inlink->format,
147 outlink->w, outlink->h, outlink->format,
148 SWS_BICUBIC, NULL, NULL, NULL);
151 av_frame_free(&fake_in);
156 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
158 AVFilterContext *context = inlink->dst;
159 SRContext *ctx = context->priv;
160 AVFilterLink *outlink = context->outputs[0];
161 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
162 DNNReturnType dnn_result;
163 const char *model_output_name = "y";
166 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
168 return AVERROR(ENOMEM);
170 av_frame_copy_props(out, in);
172 if (ctx->sws_pre_scale) {
173 sws_scale(ctx->sws_pre_scale,
174 (const uint8_t **)in->data, in->linesize, 0, in->height,
175 out->data, out->linesize);
176 dnn_result = (ctx->dnn_module->execute_model)(ctx->model, "x", out,
177 (const char **)&model_output_name, 1, out);
179 dnn_result = (ctx->dnn_module->execute_model)(ctx->model, "x", in,
180 (const char **)&model_output_name, 1, out);
183 if (dnn_result != DNN_SUCCESS){
184 av_log(ctx, AV_LOG_ERROR, "failed to execute loaded model\n");
190 if (ctx->sws_uv_scale) {
191 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 1), in->linesize + 1,
192 0, ctx->sws_uv_height, out->data + 1, out->linesize + 1);
193 sws_scale(ctx->sws_uv_scale, (const uint8_t **)(in->data + 2), in->linesize + 2,
194 0, ctx->sws_uv_height, out->data + 2, out->linesize + 2);
198 return ff_filter_frame(outlink, out);
201 static av_cold void uninit(AVFilterContext *context)
203 SRContext *sr_context = context->priv;
205 if (sr_context->dnn_module){
206 (sr_context->dnn_module->free_model)(&sr_context->model);
207 av_freep(&sr_context->dnn_module);
210 sws_freeContext(sr_context->sws_uv_scale);
211 sws_freeContext(sr_context->sws_pre_scale);
214 static const AVFilterPad sr_inputs[] = {
217 .type = AVMEDIA_TYPE_VIDEO,
218 .filter_frame = filter_frame,
223 static const AVFilterPad sr_outputs[] = {
226 .config_props = config_output,
227 .type = AVMEDIA_TYPE_VIDEO,
232 AVFilter ff_vf_sr = {
234 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
235 .priv_size = sizeof(SRContext),
238 .query_formats = query_formats,
240 .outputs = sr_outputs,
241 .priv_class = &sr_class,