BURNED-IN CAPTION TEXT DETECTION

1 . A method comprising:

inputting a frame sample of a video into a prediction network of a discriminator;

analyzing the frame sample to determine whether the frame sample includes burned-in caption text;

when the frame sample is determined to include burned-in caption text:

sending the frame to a recognition engine to perform a recognition process on the frame sample;

performing the recognition process on the frame sample to recognize text in the frame sample;

outputting the text for a service to be performed for the video; and

when the frame sample is determined to not include burned-in caption text, bypassing the recognition engine and not performing the recognition process on the frame sample.

2 . The method of claim 1 , wherein when the frame includes non-caption text, the prediction network determines the frame sample does not include burned-in caption text.

3 . The method of claim 1 , further comprising:

receiving frames of the video; and

analyzing the frames of the video to select a first portion of the frames for input into the prediction network, wherein a second portion of the frames is not input into the prediction network or the recognition engine.

4 . The method of claim 3 , wherein the first portion of the frames is selected based on a time-based sampling that selects frames based on a time associated with the frames.

5 . The method of claim 1 , further comprising:

receiving the frame of the video; and

selecting a first portion of the frame for input into the prediction network, wherein a second portion of the frame is not input into the prediction network or the recognition engine.

6 . The method of claim 5 , wherein the first portion of frame is selected based on an area that is designated as likely to include burned-in caption text.

7 . The method of claim 1 , wherein the burned-in caption text is inserted in the frame before encoding of the frame.

8 . The method of claim 1 , wherein:

determining whether the frame include burned-in caption text comprises not recognizing non-caption text as burned-in caption text, and

the non-caption text is inserted in the frame after encoding of the frame or captured by a camera.

9 . The method of claim 1 , further comprising:

training the prediction network to recognize patterns in frames for burned-in caption text, wherein parameters for the prediction network are adjusted to distinguish between burned-in caption text and non-caption text.

10 . The method of claim 9 , further comprising:

training the prediction network to recognize patterns in frames for non-caption text, wherein parameters for the prediction network are adjusted to distinguish between burned-in caption text and non-caption text.

11 . The method of claim 1 , wherein training the prediction network comprises:

labeling training frames with a label that the frame includes burned-in caption text or does not include burned-in caption text;

analyzing the frames with the prediction network to output a score of whether the frames include burned-in caption text; and

adjusting parameters of the prediction network based on a comparison of labels of the frames and the respective score for the frames.

12 . The method of claim 1 , wherein analyzing the frame sample comprises:

analyzing pixels of the frame to determine a score of a probability that the frame includes burned-in caption text; and

comparing the score to a threshold to determine whether to select the frame for input into the recognition engine.

13 . The method of claim 1 , wherein analyzing the frame sample comprises:

analyzing pixels of the frame to determine a first score of a probability that the frame includes burned-in caption text;

analyzing pixels of the frame to determine a second score of a probability that the frame does not include burned-in caption text;

comparing the first score to a first threshold to determine whether to select the frame for input into the recognition engine; and

comparing the second score to a second threshold to determine whether to select the frame for bypass of the recognition engine.

14 . The method of claim 1 , wherein analyzing the frame sample comprises:

when the frame sample is determined to include non-caption text, bypassing the recognition engine.

15 . The method of claim 1 , further comprising:

analyzing the text to determine a language of the text; and

performing a service for the video based on the language.

16 . The method of claim 1 , wherein performing the service comprises:

translating the text to another language.

17 . A non-transitory computer-readable storage medium having stored thereon computer executable instructions, which when executed by a computing device, cause the computing device to be operable for:

inputting a frame sample of a video into a prediction network of a discriminator;

analyzing the frame sample to determine whether the frame sample includes burned-in caption text;

when the frame sample is determined to include burned-in caption text:

sending the frame to a recognition engine to perform a recognition process on the frame sample;

performing the recognition process on the frame sample to recognize text in the frame sample;

outputting the text for a service to be performed for the video; and

when the frame sample is determined to not include burned-in caption text, bypassing the recognition engine and not performing the recognition process on the frame sample.

18 . A method comprising:

inputting frame samples of a video into a prediction network of a discriminator to generate a score for whether the frame samples include burned-in caption text;

training the prediction network to recognize patterns in frames for burned-in caption text;

adjusting parameters for the prediction network to adjust the score for frames with burned-in caption text to indicate respective frames include burned-in caption text;

training the prediction network to recognize patterns in frames for non-caption text; and

adjusting parameters for the prediction network to adjust the score for frames with non-caption text to indicate respective frames do not include burned-in caption text, wherein prediction network is trained to output a score a frame sample that includes non-caption text does not include burned-in caption text.

19 . The method of claim 18 , further operable to:

using the prediction network to determine whether a frame sample includes burned-in caption text or does not include burned-in caption text.

20 . The method of claim 18 , wherein:

the prediction network outputs a score that indicates the frame sample includes burned-in caption text when the frame sample includes burned-in caption text, and

the prediction network outputs a score that indicates the frame sample does not include burned-in caption text when the frame sample includes non-caption text.