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;
120 sr_context->input.width = inlink->w * sr_context->scale_factor;
121 sr_context->input.height = inlink->h * sr_context->scale_factor;
122 sr_context->input.channels = 1;
124 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, &sr_context->output);
125 if (result != DNN_SUCCESS){
126 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
130 if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
131 sr_context->input.width = inlink->w;
132 sr_context->input.height = inlink->h;
133 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, &sr_context->output);
134 if (result != DNN_SUCCESS){
135 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
138 sr_context->scale_factor = 0;
140 outlink->h = sr_context->output.height;
141 outlink->w = sr_context->output.width;
142 sr_context->sws_contexts[1] = sws_getContext(sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAY8,
143 sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAYF32,
144 0, NULL, NULL, NULL);
145 sr_context->sws_input_linesize = sr_context->input.width << 2;
146 sr_context->sws_contexts[2] = sws_getContext(sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAYF32,
147 sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAY8,
148 0, NULL, NULL, NULL);
149 sr_context->sws_output_linesize = sr_context->output.width << 2;
150 if (!sr_context->sws_contexts[1] || !sr_context->sws_contexts[2]){
151 av_log(context, AV_LOG_ERROR, "could not create SwsContext for conversions\n");
152 return AVERROR(ENOMEM);
154 if (sr_context->scale_factor){
155 sr_context->sws_contexts[0] = sws_getContext(inlink->w, inlink->h, inlink->format,
156 outlink->w, outlink->h, outlink->format,
157 SWS_BICUBIC, NULL, NULL, NULL);
158 if (!sr_context->sws_contexts[0]){
159 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
160 return AVERROR(ENOMEM);
162 sr_context->sws_slice_h = inlink->h;
164 if (inlink->format != AV_PIX_FMT_GRAY8){
165 sws_src_h = sr_context->input.height;
166 sws_src_w = sr_context->input.width;
167 sws_dst_h = sr_context->output.height;
168 sws_dst_w = sr_context->output.width;
170 switch (inlink->format){
171 case AV_PIX_FMT_YUV420P:
172 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 1);
173 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
174 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 1);
175 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
177 case AV_PIX_FMT_YUV422P:
178 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 1);
179 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 1);
181 case AV_PIX_FMT_YUV444P:
183 case AV_PIX_FMT_YUV410P:
184 sws_src_h = AV_CEIL_RSHIFT(sws_src_h, 2);
185 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
186 sws_dst_h = AV_CEIL_RSHIFT(sws_dst_h, 2);
187 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
189 case AV_PIX_FMT_YUV411P:
190 sws_src_w = AV_CEIL_RSHIFT(sws_src_w, 2);
191 sws_dst_w = AV_CEIL_RSHIFT(sws_dst_w, 2);
194 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling for given input pixel format");
197 sr_context->sws_contexts[0] = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
198 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
199 SWS_BICUBIC, NULL, NULL, NULL);
200 if (!sr_context->sws_contexts[0]){
201 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
202 return AVERROR(ENOMEM);
204 sr_context->sws_slice_h = sws_src_h;
211 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
213 AVFilterContext *context = inlink->dst;
214 SRContext *sr_context = context->priv;
215 AVFilterLink *outlink = context->outputs[0];
216 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
217 DNNReturnType dnn_result;
220 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
222 return AVERROR(ENOMEM);
224 av_frame_copy_props(out, in);
225 out->height = sr_context->output.height;
226 out->width = sr_context->output.width;
227 if (sr_context->scale_factor){
228 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)in->data, in->linesize,
229 0, sr_context->sws_slice_h, out->data, out->linesize);
231 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)out->data, out->linesize,
232 0, out->height, (uint8_t * const*)(&sr_context->input.data),
233 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
235 if (sr_context->sws_contexts[0]){
236 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 1), in->linesize + 1,
237 0, sr_context->sws_slice_h, out->data + 1, out->linesize + 1);
238 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 2), in->linesize + 2,
239 0, sr_context->sws_slice_h, out->data + 2, out->linesize + 2);
242 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)in->data, in->linesize,
243 0, in->height, (uint8_t * const*)(&sr_context->input.data),
244 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
248 dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model);
249 if (dnn_result != DNN_SUCCESS){
250 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
254 sws_scale(sr_context->sws_contexts[2], (const uint8_t *[4]){(const uint8_t *)sr_context->output.data, 0, 0, 0},
255 (const int[4]){sr_context->sws_output_linesize, 0, 0, 0},
256 0, out->height, (uint8_t * const*)out->data, out->linesize);
258 return ff_filter_frame(outlink, out);
261 static av_cold void uninit(AVFilterContext *context)
264 SRContext *sr_context = context->priv;
266 if (sr_context->dnn_module){
267 (sr_context->dnn_module->free_model)(&sr_context->model);
268 av_freep(&sr_context->dnn_module);
271 for (i = 0; i < 3; ++i){
272 if (sr_context->sws_contexts[i]){
273 sws_freeContext(sr_context->sws_contexts[i]);
278 static const AVFilterPad sr_inputs[] = {
281 .type = AVMEDIA_TYPE_VIDEO,
282 .config_props = config_props,
283 .filter_frame = filter_frame,
288 static const AVFilterPad sr_outputs[] = {
291 .type = AVMEDIA_TYPE_VIDEO,
296 AVFilter ff_vf_sr = {
298 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
299 .priv_size = sizeof(SRContext),
302 .query_formats = query_formats,
304 .outputs = sr_outputs,
305 .priv_class = &sr_class,
306 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC | AVFILTER_FLAG_SLICE_THREADS,