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;
47 struct SwsContext *sws_contexts[3];
48 int sws_slice_h, sws_input_linesize, sws_output_linesize;
51 #define OFFSET(x) offsetof(SRContext, x)
52 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
53 static const AVOption sr_options[] = {
54 { "dnn_backend", "DNN backend used for model execution", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS, "backend" },
55 { "native", "native backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, 0, 0, FLAGS, "backend" },
56 #if (CONFIG_LIBTENSORFLOW == 1)
57 { "tensorflow", "tensorflow backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, 0, 0, FLAGS, "backend" },
59 { "scale_factor", "scale factor for SRCNN model", OFFSET(scale_factor), AV_OPT_TYPE_INT, { .i64 = 2 }, 2, 4, FLAGS },
60 { "model", "path to model file specifying network architecture and its parameters", OFFSET(model_filename), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
64 AVFILTER_DEFINE_CLASS(sr);
66 static av_cold int init(AVFilterContext *context)
68 SRContext *sr_context = context->priv;
70 sr_context->dnn_module = ff_get_dnn_module(sr_context->backend_type);
71 if (!sr_context->dnn_module){
72 av_log(context, AV_LOG_ERROR, "could not create DNN module for requested backend\n");
73 return AVERROR(ENOMEM);
76 if (!sr_context->model_filename){
77 av_log(context, AV_LOG_ERROR, "model file for network was not specified\n");
80 if (!sr_context->dnn_module->load_model) {
81 av_log(context, AV_LOG_ERROR, "load_model for network was not specified\n");
84 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->input.dt = DNN_FLOAT;
91 sr_context->sws_contexts[0] = NULL;
92 sr_context->sws_contexts[1] = NULL;
93 sr_context->sws_contexts[2] = NULL;
98 static int query_formats(AVFilterContext *context)
100 const enum AVPixelFormat pixel_formats[] = {AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
101 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_GRAY8,
103 AVFilterFormats *formats_list;
105 formats_list = ff_make_format_list(pixel_formats);
107 av_log(context, AV_LOG_ERROR, "could not create formats list\n");
108 return AVERROR(ENOMEM);
111 return ff_set_common_formats(context, formats_list);
114 static int config_props(AVFilterLink *inlink)
116 AVFilterContext *context = inlink->dst;
117 SRContext *sr_context = context->priv;
118 AVFilterLink *outlink = context->outputs[0];
119 DNNReturnType result;
120 int sws_src_h, sws_src_w, sws_dst_h, sws_dst_w;
121 const char *model_output_name = "y";
123 sr_context->input.width = inlink->w * sr_context->scale_factor;
124 sr_context->input.height = inlink->h * sr_context->scale_factor;
125 sr_context->input.channels = 1;
127 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
128 if (result != DNN_SUCCESS){
129 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
133 result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
134 if (result != DNN_SUCCESS){
135 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
139 if (sr_context->input.height != sr_context->output.height || sr_context->input.width != sr_context->output.width){
140 sr_context->input.width = inlink->w;
141 sr_context->input.height = inlink->h;
142 result = (sr_context->model->set_input_output)(sr_context->model->model, &sr_context->input, "x", &model_output_name, 1);
143 if (result != DNN_SUCCESS){
144 av_log(context, AV_LOG_ERROR, "could not set input and output for the model\n");
147 result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
148 if (result != DNN_SUCCESS){
149 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
152 sr_context->scale_factor = 0;
154 outlink->h = sr_context->output.height;
155 outlink->w = sr_context->output.width;
156 sr_context->sws_contexts[1] = sws_getContext(sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAY8,
157 sr_context->input.width, sr_context->input.height, AV_PIX_FMT_GRAYF32,
158 0, NULL, NULL, NULL);
159 sr_context->sws_input_linesize = sr_context->input.width << 2;
160 sr_context->sws_contexts[2] = sws_getContext(sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAYF32,
161 sr_context->output.width, sr_context->output.height, AV_PIX_FMT_GRAY8,
162 0, NULL, NULL, NULL);
163 sr_context->sws_output_linesize = sr_context->output.width << 2;
164 if (!sr_context->sws_contexts[1] || !sr_context->sws_contexts[2]){
165 av_log(context, AV_LOG_ERROR, "could not create SwsContext for conversions\n");
166 return AVERROR(ENOMEM);
168 if (sr_context->scale_factor){
169 sr_context->sws_contexts[0] = sws_getContext(inlink->w, inlink->h, inlink->format,
170 outlink->w, outlink->h, outlink->format,
171 SWS_BICUBIC, NULL, NULL, NULL);
172 if (!sr_context->sws_contexts[0]){
173 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
174 return AVERROR(ENOMEM);
176 sr_context->sws_slice_h = inlink->h;
178 if (inlink->format != AV_PIX_FMT_GRAY8){
179 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
180 sws_src_h = AV_CEIL_RSHIFT(sr_context->input.height, desc->log2_chroma_h);
181 sws_src_w = AV_CEIL_RSHIFT(sr_context->input.width, desc->log2_chroma_w);
182 sws_dst_h = AV_CEIL_RSHIFT(sr_context->output.height, desc->log2_chroma_h);
183 sws_dst_w = AV_CEIL_RSHIFT(sr_context->output.width, desc->log2_chroma_w);
185 sr_context->sws_contexts[0] = sws_getContext(sws_src_w, sws_src_h, AV_PIX_FMT_GRAY8,
186 sws_dst_w, sws_dst_h, AV_PIX_FMT_GRAY8,
187 SWS_BICUBIC, NULL, NULL, NULL);
188 if (!sr_context->sws_contexts[0]){
189 av_log(context, AV_LOG_ERROR, "could not create SwsContext for scaling\n");
190 return AVERROR(ENOMEM);
192 sr_context->sws_slice_h = sws_src_h;
199 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
201 AVFilterContext *context = inlink->dst;
202 SRContext *sr_context = context->priv;
203 AVFilterLink *outlink = context->outputs[0];
204 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
205 DNNReturnType dnn_result;
208 av_log(context, AV_LOG_ERROR, "could not allocate memory for output frame\n");
210 return AVERROR(ENOMEM);
212 av_frame_copy_props(out, in);
213 out->height = sr_context->output.height;
214 out->width = sr_context->output.width;
215 if (sr_context->scale_factor){
216 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)in->data, in->linesize,
217 0, sr_context->sws_slice_h, out->data, out->linesize);
219 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)out->data, out->linesize,
220 0, out->height, (uint8_t * const*)(&sr_context->input.data),
221 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
223 if (sr_context->sws_contexts[0]){
224 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 1), in->linesize + 1,
225 0, sr_context->sws_slice_h, out->data + 1, out->linesize + 1);
226 sws_scale(sr_context->sws_contexts[0], (const uint8_t **)(in->data + 2), in->linesize + 2,
227 0, sr_context->sws_slice_h, out->data + 2, out->linesize + 2);
230 sws_scale(sr_context->sws_contexts[1], (const uint8_t **)in->data, in->linesize,
231 0, in->height, (uint8_t * const*)(&sr_context->input.data),
232 (const int [4]){sr_context->sws_input_linesize, 0, 0, 0});
236 dnn_result = (sr_context->dnn_module->execute_model)(sr_context->model, &sr_context->output, 1);
237 if (dnn_result != DNN_SUCCESS){
238 av_log(context, AV_LOG_ERROR, "failed to execute loaded model\n");
242 sws_scale(sr_context->sws_contexts[2], (const uint8_t *[4]){(const uint8_t *)sr_context->output.data, 0, 0, 0},
243 (const int[4]){sr_context->sws_output_linesize, 0, 0, 0},
244 0, out->height, (uint8_t * const*)out->data, out->linesize);
246 return ff_filter_frame(outlink, out);
249 static av_cold void uninit(AVFilterContext *context)
252 SRContext *sr_context = context->priv;
254 if (sr_context->dnn_module){
255 (sr_context->dnn_module->free_model)(&sr_context->model);
256 av_freep(&sr_context->dnn_module);
259 for (i = 0; i < 3; ++i){
260 sws_freeContext(sr_context->sws_contexts[i]);
264 static const AVFilterPad sr_inputs[] = {
267 .type = AVMEDIA_TYPE_VIDEO,
268 .config_props = config_props,
269 .filter_frame = filter_frame,
274 static const AVFilterPad sr_outputs[] = {
277 .type = AVMEDIA_TYPE_VIDEO,
282 AVFilter ff_vf_sr = {
284 .description = NULL_IF_CONFIG_SMALL("Apply DNN-based image super resolution to the input."),
285 .priv_size = sizeof(SRContext),
288 .query_formats = query_formats,
290 .outputs = sr_outputs,
291 .priv_class = &sr_class,