Bitrate video transcoding based on video coding complexity estimation

A computer system for adaptively transcoding a source video of a video hosting service based on video coding complexity, the system comprising:

• a video coding complexity model training module configured to train a video coding complexity model with a plurality of videos selected from a video corpus of the video hosting service, the video coding complexity model providing a video coding complexity score and trained using linear regression to model correlations between features extracted from the selected plurality of videos and normalized bitrates of the selected plurality of videos;
• a target bitrate estimation module configured to:
• receive the source video having a plurality of coding parameters and a target frame rate and target resolution;
• determine a set of sequence-level features from the source video, the features in the set of sequence-level features selected based on the modeled correlations;
• normalize the sequence-level features in the set;
• apply the trained video coding complexity model to the source video to obtain a video coding complexity score of the source video, wherein the video coding complexity score of the source video is generated by the trained video coding complexity model based upon the set of normalized sequence-level features;
• and
• apply a trained rate-distortion model to the source video to calculate an initial target bitrate of the source video based on the video coding complexity score;
• a bitrate refinement module configured to:
• apply a trained scaling model to adjust the initial target bitrate based on at least one of the target frame rate and target resolution of the source video, the bitrate refinement module coupled to the target bitrate estimation module;
• and
• an adaptive video encoder configured to encode the source video with the adjusted target bitrate, the adaptive video encoder coupled to the bitrate refinement module.

A computer program product having a non-transitory computer-readable storage medium having executable computer program instructions recorded thereon for adaptively transcoding a source video of a video hosting service based on video coding complexity, the computer program instructions configuring a computer system to comprise:

• a video coding complexity model training module configured to train a video coding complexity model with a plurality of videos selected from a video corpus of the video hosting service, the video coding complexity model providing a video coding complexity score and trained using linear regression to model correlations between features extracted from the selected plurality of videos and normalized bitrates of the selected plurality of videos;
• a target bitrate estimation module configured to:
• receive the source video having a plurality of coding parameters and a target frame rate and target resolution;
• determine a set of sequence-level features from the source video, the features in the set of sequence-level features selected based on the modeled correlations;
• normalize the sequence-level features in the set;
• apply the trained video coding complexity model to the source video to obtain a video coding complexity score of the source video, wherein the video coding complexity score of the source video is generated by the trained video coding complexity model based upon the set of normalized sequence-level features;
• and
• apply a trained rate-distortion model to the source video to calculate an initial target bitrate of the source video based on the video coding complexity score;
• a bitrate refinement module configured to:
• apply a trained scaling model to adjust the initial target bitrate based on at least one of the target frame rate and target resolution of the source video, the bitrate refinement module coupled to the target bitrate estimation module;
• and
• an adaptive video encoder configured to encode the source video with the adjusted target bitrate, the adaptive video encoder coupled to the bitrate refinement module.

A method of training a rate-distortion model and a scaling model for encoding videos, the method comprising:

• training a video coding complexity model with a plurality of videos in multiple video formats selected from a video corpus of a video hosting service, the video coding complexity model providing a video coding complexity score and trained using linear regression to model correlations between features extracted from the selected plurality of videos and normalized bitrates of the selected plurality of videos;
• encoding the selected videos in multiple video formats;
• obtaining a plurality of rate-distortion data from the encoded videos;
• and
• for each of the selected videos:
• determining a set of sequence-level features from the video, the features in the set of sequence-level features selected based on the modeled correlations;
• normalizing the sequence-level features in the set;
• and
• applying the trained video coding complexity model to the video to obtain a video coding complexity score for the selected video, wherein the video coding complexity score is generated by the trained video coding complexity model based upon the set of normalized sequence-level features;
• estimating a rate-distortion model based upon video coding complexity scores of the selected videos;
• and
• estimating the scaling model based on the obtained plurality of rate-distortion data.

A computer system for training a rate-distortion model and a scaling model for encoding videos, comprising:

• a video coding complexity training module configured to train a video coding complexity model with a plurality of videos in multiple video formats selected from a video corpus of a video hosting service, the video coding complexity model providing a video coding complexity score and trained using linear regression to model correlations between features extracted from the selected plurality of videos and normalized bitrates of the selected plurality of videos,
• wherein the video coding complexity training module is further configured to, for each of the selected videos:
• determine a set of sequence-level features from the video, the features in the set of sequence-level features selected based on the modeled correlations;
• normalize the sequence-level features in the set;
• and
• generate a video coding complexity score for the video based upon the set of normalized sequence-level features;
• a video encoder configured to encode the selected videos in multiple video formats;
• a rate-distortion manager configured to obtain a plurality of rate-distortion data from the encoded videos;
• a rate-distortion estimation module configured to estimate a rate-distortion model based upon video coding complexity scores of the selected videos;
• and
• a scaling model estimation module configured to estimate the scaling model based on the obtained plurality of rate-distortion data.

A computer program product having a non-transitory computer-readable storage medium having executable computer program instructions recorded thereon for training a rate-distortion model and a scaling model, the computer program instructions configuring a computer system to comprise:

• a video coding complexity training module configured to train a video coding complexity model with a plurality of videos in multiple video formats selected from a video corpus of a video hosting service, the video coding complexity model trained using linear regression to model correlations between features extracted from the selected plurality of videos and normalized bitrates of the selected set of videos,
• wherein the video coding complexity training module is further configured to, for each of the selected videos:
• determine a set of sequence-level features from the video, the features in the set of sequence-level features selected based on the modeled correlations;
• normalize the sequence-level features in the set;
• and
• generate a video coding complexity score for the video based upon the set of normalized sequence-level features;
• a video encoder configured to encode the selected videos in multiple video formats;
• a rate-distortion manager configured to obtain a plurality of rate-distortion data from the encoded videos;
• a rate-distortion estimation module configured to estimate a rate-distortion model based upon video coding complexity scores of the selected videos;
• and
• a scaling model estimation module configured to estimate the scaling model based on the obtained plurality of rate-distortion data.