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");
77 if (!sr_context->dnn_module->load_model) {
78 av_log(context, AV_LOG_ERROR, "load_model for network was not specified\n");
81 sr_context->model = (sr_context->dnn_module->load_model)(sr_context->model_filename);
84 if (!sr_context->model){
85 av_log(context, AV_LOG_ERROR, "could not load DNN model\n");
89 sr_context->sws_contexts[0] = NULL;
90 sr_context->sws_contexts[1] = NULL;
91 sr_context->sws_contexts[2] = NULL;
96 static int query_formats(AVFilterContext *context)
98 const enum AVPixelFormat pixel_formats[] = {AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
99 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_GRAY8,
101 AVFilterFormats *formats_list;
103 formats_list = ff_make_format_list(pixel_formats);
105 av_log(context, AV_LOG_ERROR, "could not create formats list\n");
106 return AVERROR(ENOMEM);
109 return ff_set_common_formats(context, formats_list);
112 static int config_props(AVFilterLink *inlink)
114 AVFilterContext *context = inlink->dst;
115 SRContext *sr_context = context->priv;
116 AVFilterLink *outlink = context->outputs[0];
117 DNNReturnType result;
118 int sws_src_h, sws_src_w, sws_dst_h, sws_dst_w;
119 const char *model_output_name = "y";
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, "x", &model_output_name, 1);
126 if (result != DNN_SUCCESS){
127 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
131 result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
132 if (result != DNN_SUCCESS){
133 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
137 if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
138 sr_context->input.width = inlink->w;
139 sr_context->input.height = inlink->h;
140 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
141 if (result != DNN_SUCCESS){
142 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
145 result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
146 if (result != DNN_SUCCESS){
147 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
150 sr_context->scale_factor = 0;
152 outlink->h = sr_context->output.height;
153 outlink->w = sr_context->output.width;
154 sr_context->sws_contexts[1] = sws_getContext(sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAY8,
155 sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAYF32,
156 0, NULL, NULL, NULL);
157 sr_context->sws_input_linesize = sr_context->input.width << 2;
158 sr_context->sws_contexts[2] = sws_getContext(sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAYF32,
159 sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAY8,
160 0, NULL, NULL, NULL);
161 sr_context->sws_output_linesize = sr_context->output.width << 2;
162 if (!sr_context->sws_contexts[1] || !sr_context->sws_contexts[2]){
163 av_log(context, AV_LOG_ERROR, "could not create SwsContext for conversions\n");
164 return AVERROR(ENOMEM);
166 if (sr_context->scale_factor){
167 sr_context->sws_contexts[0] = sws_getContext(inlink->w, inlink->h, inlink->format,
168 outlink->w, outlink->h, outlink->format,
169 SWS_BICUBIC, NULL, NULL, NULL);
170 if (!sr_context->sws_contexts[0]){
171 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
172 return AVERROR(ENOMEM);
174 sr_context->sws_slice_h = inlink->h;
176 if (inlink->format != AV_PIX_FMT_GRAY8){
177 sws_src_h = sr_context->input.height;
178 sws_src_w = sr_context->input.width;
179 sws_dst_h = sr_context->output.height;
180 sws_dst_w = sr_context->output.width;
182 switch (inlink->format){
183 case AV_PIX_FMT_YUV420P:
184 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 1);
185 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
186 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 1);
187 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
189 case AV_PIX_FMT_YUV422P:
190 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
191 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
193 case AV_PIX_FMT_YUV444P:
195 case AV_PIX_FMT_YUV410P:
196 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 2);
197 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
198 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 2);
199 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
201 case AV_PIX_FMT_YUV411P:
202 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
203 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
206 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling for given input pixel format");
209 sr_context->sws_contexts[0] = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
210 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
211 SWS_BICUBIC, NULL, NULL, NULL);
212 if (!sr_context->sws_contexts[0]){
213 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
214 return AVERROR(ENOMEM);
216 sr_context->sws_slice_h = sws_src_h;
223 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
225 AVFilterContext *context = inlink->dst;
226 SRContext *sr_context = context->priv;
227 AVFilterLink *outlink = context->outputs[0];
228 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
229 DNNReturnType dnn_result;
232 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
234 return AVERROR(ENOMEM);
236 av_frame_copy_props(out, in);
237 out->height = sr_context->output.height;
238 out->width = sr_context->output.width;
239 if (sr_context->scale_factor){
240 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)in->data, in->linesize,
241 0, sr_context->sws_slice_h, out->data, out->linesize);
243 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)out->data, out->linesize,
244 0, out->height, (uint8_t * const*)(&sr_context->input.data),
245 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
247 if (sr_context->sws_contexts[0]){
248 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 1), in->linesize + 1,
249 0, sr_context->sws_slice_h, out->data + 1, out->linesize + 1);
250 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 2), in->linesize + 2,
251 0, sr_context->sws_slice_h, out->data + 2, out->linesize + 2);
254 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)in->data, in->linesize,
255 0, in->height, (uint8_t * const*)(&sr_context->input.data),
256 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
260 dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
261 if (dnn_result != DNN_SUCCESS){
262 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
266 sws_scale(sr_context->sws_contexts[2], (const uint8_t *[4]){(const uint8_t *)sr_context->output.data, 0, 0, 0},
267 (const int[4]){sr_context->sws_output_linesize, 0, 0, 0},
268 0, out->height, (uint8_t * const*)out->data, out->linesize);
270 return ff_filter_frame(outlink, out);
273 static av_cold void uninit(AVFilterContext *context)
276 SRContext *sr_context = context->priv;
278 if (sr_context->dnn_module){
279 (sr_context->dnn_module->free_model)(&sr_context->model);
280 av_freep(&sr_context->dnn_module);
283 for (i = 0; i < 3; ++i){
284 if (sr_context->sws_contexts[i]){
285 sws_freeContext(sr_context->sws_contexts[i]);
290 static const AVFilterPad sr_inputs[] = {
293 .type = AVMEDIA_TYPE_VIDEO,
294 .config_props = config_props,
295 .filter_frame = filter_frame,
300 static const AVFilterPad sr_outputs[] = {
303 .type = AVMEDIA_TYPE_VIDEO,
308 AVFilter ff_vf_sr = {
310 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
311 .priv_size = sizeof(SRContext),
314 .query_formats = query_formats,
316 .outputs = sr_outputs,
317 .priv_class = &sr_class,
318 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,