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 "libavformat/avio.h"
33 #include "libswscale/swscale.h"
34 #include "dnn_interface.h"
36 typedef struct SRContext {
40 DNNBackendType backend_type;
41 DNNModule *dnn_module;
43 DNNData input, output;
45 struct SwsContext *sws_contexts[3];
46 int sws_slice_h, sws_input_linesize, sws_output_linesize;
49 #define OFFSET(x) offsetof(SRContext, x)
50 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
51 static const AVOption sr_options[] = {
52 { "dnn_backend", "DNN backend used for model execution", OFFSET(backend_type), AV_OPT_TYPE_FLAGS, { .i64 = 0 }, 0, 1, FLAGS, "backend" },
53 { "native", "native backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, 0, 0, FLAGS, "backend" },
54 #if (CONFIG_LIBTENSORFLOW == 1)
55 { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, "backend" },
57 { "scale_factor", "scale factor for SRCNN model", OFFSET(scale_factor), AV_OPT_TYPE_INT, { .i64 = 2 }, 2, 4, FLAGS },
58 { "model", "path to model file specifying network architecture and its parameters", OFFSET(model_filename), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
62 AVFILTER_DEFINE_CLASS(sr);
64 static av_cold int init(AVFilterContext *context)
66 SRContext *sr_context = context->priv;
68 sr_context->dnn_module = ff_get_dnn_module(sr_context->backend_type);
69 if (!sr_context->dnn_module){
70 av_log(context, AV_LOG_ERROR, "could not create DNN module for requested backend\n");
71 return AVERROR(ENOMEM);
73 if (!sr_context->model_filename){
74 av_log(context, AV_LOG_ERROR, "model file for network was not specified\n");
78 if (!sr_context->dnn_module->load_model) {
79 av_log(context, AV_LOG_ERROR, "load_model for network was not specified\n");
82 sr_context->model = (sr_context->dnn_module->load_model)(sr_context->model_filename);
85 if (!sr_context->model){
86 av_log(context, AV_LOG_ERROR, "could not load DNN model\n");
90 sr_context->sws_contexts[0] = NULL;
91 sr_context->sws_contexts[1] = NULL;
92 sr_context->sws_contexts[2] = NULL;
97 static int query_formats(AVFilterContext *context)
99 const enum AVPixelFormat pixel_formats[] = {AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
100 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_GRAY8,
102 AVFilterFormats *formats_list;
104 formats_list = ff_make_format_list(pixel_formats);
106 av_log(context, AV_LOG_ERROR, "could not create formats list\n");
107 return AVERROR(ENOMEM);
110 return ff_set_common_formats(context, formats_list);
113 static int config_props(AVFilterLink *inlink)
115 AVFilterContext *context = inlink->dst;
116 SRContext *sr_context = context->priv;
117 AVFilterLink *outlink = context->outputs[0];
118 DNNReturnType result;
119 int sws_src_h, sws_src_w, sws_dst_h, sws_dst_w;
121 sr_context->input.width = inlink->w * sr_context->scale_factor;
122 sr_context->input.height = inlink->h * sr_context->scale_factor;
123 sr_context->input.channels = 1;
125 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, &sr_context->output);
126 if (result != DNN_SUCCESS){
127 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
131 if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
132 sr_context->input.width = inlink->w;
133 sr_context->input.height = inlink->h;
134 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, &sr_context->output);
135 if (result != DNN_SUCCESS){
136 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
139 sr_context->scale_factor = 0;
141 outlink->h = sr_context->output.height;
142 outlink->w = sr_context->output.width;
143 sr_context->sws_contexts[1] = sws_getContext(sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAY8,
144 sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAYF32,
145 0, NULL, NULL, NULL);
146 sr_context->sws_input_linesize = sr_context->input.width << 2;
147 sr_context->sws_contexts[2] = sws_getContext(sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAYF32,
148 sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAY8,
149 0, NULL, NULL, NULL);
150 sr_context->sws_output_linesize = sr_context->output.width << 2;
151 if (!sr_context->sws_contexts[1] || !sr_context->sws_contexts[2]){
152 av_log(context, AV_LOG_ERROR, "could not create SwsContext for conversions\n");
153 return AVERROR(ENOMEM);
155 if (sr_context->scale_factor){
156 sr_context->sws_contexts[0] = sws_getContext(inlink->w, inlink->h, inlink->format,
157 outlink->w, outlink->h, outlink->format,
158 SWS_BICUBIC, NULL, NULL, NULL);
159 if (!sr_context->sws_contexts[0]){
160 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
161 return AVERROR(ENOMEM);
163 sr_context->sws_slice_h = inlink->h;
166 if (inlink->format != AV_PIX_FMT_GRAY8){
167 sws_src_h = sr_context->input.height;
168 sws_src_w = sr_context->input.width;
169 sws_dst_h = sr_context->output.height;
170 sws_dst_w = sr_context->output.width;
172 switch (inlink->format){
173 case AV_PIX_FMT_YUV420P:
174 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 1);
175 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
176 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 1);
177 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
179 case AV_PIX_FMT_YUV422P:
180 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
181 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
183 case AV_PIX_FMT_YUV444P:
185 case AV_PIX_FMT_YUV410P:
186 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 2);
187 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
188 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 2);
189 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
191 case AV_PIX_FMT_YUV411P:
192 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
193 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
196 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling for given input pixel format");
199 sr_context->sws_contexts[0] = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
200 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
201 SWS_BICUBIC, NULL, NULL, NULL);
202 if (!sr_context->sws_contexts[0]){
203 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
204 return AVERROR(ENOMEM);
206 sr_context->sws_slice_h = sws_src_h;
214 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
216 AVFilterContext *context = inlink->dst;
217 SRContext *sr_context = context->priv;
218 AVFilterLink *outlink = context->outputs[0];
219 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
220 DNNReturnType dnn_result;
223 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
225 return AVERROR(ENOMEM);
227 av_frame_copy_props(out, in);
228 out->height = sr_context->output.height;
229 out->width = sr_context->output.width;
230 if (sr_context->scale_factor){
231 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)in->data, in->linesize,
232 0, sr_context->sws_slice_h, out->data, out->linesize);
234 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)out->data, out->linesize,
235 0, out->height, (uint8_t * const*)(&sr_context->input.data),
236 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
239 if (sr_context->sws_contexts[0]){
240 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 1), in->linesize + 1,
241 0, sr_context->sws_slice_h, out->data + 1, out->linesize + 1);
242 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 2), in->linesize + 2,
243 0, sr_context->sws_slice_h, out->data + 2, out->linesize + 2);
246 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)in->data, in->linesize,
247 0, in->height, (uint8_t * const*)(&sr_context->input.data),
248 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
252 dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model);
253 if (dnn_result != DNN_SUCCESS){
254 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
258 sws_scale(sr_context->sws_contexts[2], (const uint8_t *[4]){(const uint8_t *)sr_context->output.data, 0, 0, 0},
259 (const int[4]){sr_context->sws_output_linesize, 0, 0, 0},
260 0, out->height, (uint8_t * const*)out->data, out->linesize);
262 return ff_filter_frame(outlink, out);
265 static av_cold void uninit(AVFilterContext *context)
268 SRContext *sr_context = context->priv;
270 if (sr_context->dnn_module){
271 (sr_context->dnn_module->free_model)(&sr_context->model);
272 av_freep(&sr_context->dnn_module);
275 for (i = 0; i < 3; ++i){
276 if (sr_context->sws_contexts[i]){
277 sws_freeContext(sr_context->sws_contexts[i]);
282 static const AVFilterPad sr_inputs[] = {
285 .type = AVMEDIA_TYPE_VIDEO,
286 .config_props = config_props,
287 .filter_frame = filter_frame,
292 static const AVFilterPad sr_outputs[] = {
295 .type = AVMEDIA_TYPE_VIDEO,
300 AVFilter ff_vf_sr = {
302 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
303 .priv_size = sizeof(SRContext),
306 .query_formats = query_formats,
308 .outputs = sr_outputs,
309 .priv_class = &sr_class,
310 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,