Machine learning for recognizing and interpreting embedded information card content

Metadata for highlights of a video stream is extracted from card images embedded in the video stream. The highlights may be segments of a video stream, such as a broadcast of a sporting event, that are of particular interest to one or more users. Card images embedded in video frames of the video stream are identified and processed to extract text. The text characters may be recognized by applying a machine-learned model trained with a set of characters extracted from card images embedded in sports television programming contents. The training set of character vectors may be pre-processed to maximize metric distance between the training set members. The text may be interpreted to obtain the metadata. The metadata may be stored in association with the portion of the video stream. The metadata may provide information regarding the highlights, and may be presented concurrently with playback of the highlights.