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 sr_context->model = (sr_context->dnn_module->load_model)(sr_context->model_filename);
80 if (!sr_context->model){
81 av_log(context, AV_LOG_ERROR, "could not load DNN model\n");
85 sr_context->sws_contexts[0] = NULL;
86 sr_context->sws_contexts[1] = NULL;
87 sr_context->sws_contexts[2] = NULL;
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_props(AVFilterLink *inlink)
110 AVFilterContext *context = inlink->dst;
111 SRContext *sr_context = context->priv;
112 AVFilterLink *outlink = context->outputs[0];
113 DNNReturnType result;
114 int sws_src_h, sws_src_w, sws_dst_h, sws_dst_w;
116 sr_context->input.width = inlink->w * sr_context->scale_factor;
117 sr_context->input.height = inlink->h * sr_context->scale_factor;
118 sr_context->input.channels = 1;
120 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, &sr_context->output);
121 if (result != DNN_SUCCESS){
122 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
126 if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
127 sr_context->input.width = inlink->w;
128 sr_context->input.height = inlink->h;
129 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, &sr_context->output);
130 if (result != DNN_SUCCESS){
131 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
134 sr_context->scale_factor = 0;
136 outlink->h = sr_context->output.height;
137 outlink->w = sr_context->output.width;
138 sr_context->sws_contexts[1] = sws_getContext(sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAY8,
139 sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAYF32,
140 0, NULL, NULL, NULL);
141 sr_context->sws_input_linesize = sr_context->input.width << 2;
142 sr_context->sws_contexts[2] = sws_getContext(sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAYF32,
143 sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAY8,
144 0, NULL, NULL, NULL);
145 sr_context->sws_output_linesize = sr_context->output.width << 2;
146 if (!sr_context->sws_contexts[1] || !sr_context->sws_contexts[2]){
147 av_log(context, AV_LOG_ERROR, "could not create SwsContext for conversions\n");
148 return AVERROR(ENOMEM);
150 if (sr_context->scale_factor){
151 sr_context->sws_contexts[0] = sws_getContext(inlink->w, inlink->h, inlink->format,
152 outlink->w, outlink->h, outlink->format,
153 SWS_BICUBIC, NULL, NULL, NULL);
154 if (!sr_context->sws_contexts[0]){
155 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
156 return AVERROR(ENOMEM);
158 sr_context->sws_slice_h = inlink->h;
161 if (inlink->format != AV_PIX_FMT_GRAY8){
162 sws_src_h = sr_context->input.height;
163 sws_src_w = sr_context->input.width;
164 sws_dst_h = sr_context->output.height;
165 sws_dst_w = sr_context->output.width;
167 switch (inlink->format){
168 case AV_PIX_FMT_YUV420P:
169 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 1);
170 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
171 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 1);
172 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
174 case AV_PIX_FMT_YUV422P:
175 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
176 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
178 case AV_PIX_FMT_YUV444P:
180 case AV_PIX_FMT_YUV410P:
181 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 2);
182 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
183 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 2);
184 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
186 case AV_PIX_FMT_YUV411P:
187 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
188 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
191 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling for given input pixel format");
194 sr_context->sws_contexts[0] = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
195 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
196 SWS_BICUBIC, NULL, NULL, NULL);
197 if (!sr_context->sws_contexts[0]){
198 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
199 return AVERROR(ENOMEM);
201 sr_context->sws_slice_h = sws_src_h;
209 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
211 AVFilterContext *context = inlink->dst;
212 SRContext *sr_context = context->priv;
213 AVFilterLink *outlink = context->outputs[0];
214 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
215 DNNReturnType dnn_result;
218 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
220 return AVERROR(ENOMEM);
222 av_frame_copy_props(out, in);
223 out->height = sr_context->output.height;
224 out->width = sr_context->output.width;
225 if (sr_context->scale_factor){
226 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)in->data, in->linesize,
227 0, sr_context->sws_slice_h, out->data, out->linesize);
229 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)out->data, out->linesize,
230 0, out->height, (uint8_t * const*)(&sr_context->input.data),
231 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
234 if (sr_context->sws_contexts[0]){
235 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 1), in->linesize + 1,
236 0, sr_context->sws_slice_h, out->data + 1, out->linesize + 1);
237 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 2), in->linesize + 2,
238 0, sr_context->sws_slice_h, out->data + 2, out->linesize + 2);
241 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)in->data, in->linesize,
242 0, in->height, (uint8_t * const*)(&sr_context->input.data),
243 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
247 dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model);
248 if (dnn_result != DNN_SUCCESS){
249 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
253 sws_scale(sr_context->sws_contexts[2], (const uint8_t **)(&sr_context->output.data),
254 (const int[4]){sr_context->sws_output_linesize, 0, 0, 0},
255 0, out->height, (uint8_t * const*)out->data, out->linesize);
257 return ff_filter_frame(outlink, out);
260 static av_cold void uninit(AVFilterContext *context)
263 SRContext *sr_context = context->priv;
265 if (sr_context->dnn_module){
266 (sr_context->dnn_module->free_model)(&sr_context->model);
267 av_freep(&sr_context->dnn_module);
270 for (i = 0; i < 3; ++i){
271 if (sr_context->sws_contexts[i]){
272 sws_freeContext(sr_context->sws_contexts[i]);
277 static const AVFilterPad sr_inputs[] = {
280 .type = AVMEDIA_TYPE_VIDEO,
281 .config_props = config_props,
282 .filter_frame = filter_frame,
287 static const AVFilterPad sr_outputs[] = {
290 .type = AVMEDIA_TYPE_VIDEO,
295 AVFilter ff_vf_sr = {
297 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
298 .priv_size = sizeof(SRContext),
301 .query_formats = query_formats,
303 .outputs = sr_outputs,
304 .priv_class = &sr_class,
305 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,