Burned-In Caption Detection With OCR Frame Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video analysis systems face challenges in distinguishing between burned-in caption text and non-caption text, leading to inefficient use of computing resources and increased false positives in optical character recognition processes.
Innovation Solution
A video analysis system employs a discriminator that uses a prediction network to differentiate between frames with burned-in caption text and those without, selectively processing only frames with burned-in caption text through an optical character recognition engine, thereby reducing unnecessary computations and improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all video frames are processed through optical character recognition, then comprehensive text detection is achieved, but computing resource usage increases significantly
Solution Approach 1:
The video processing system is segmented into two distinct pathways: a fast pathway using a trained neural network for initial classification, and a slow pathway using OCR for detailed text recognition. Only frames classified as containing burned-in captions are routed to the OCR engine, thereby segmenting the computational workload to minimize resource usage while maintaining detection comprehensiveness.
Solution Approach 2:
A preliminary classification step is performed using a trained neural network before the main OCR processing. This preliminary action filters out frames that do not contain burned-in captions, so that the resource-intensive OCR engine only processes relevant frames, reducing overall computing resource consumption.
2Measurement precision
If all video frames are processed through optical character recognition, then all text is detected, but false positives increase
Solution Approach 1:
The trained neural network performs a preliminary classification of frames to identify those containing burned-in captions before they reach the OCR engine. This preliminary filtering action reduces false positives by ensuring that only frames with actual burned-in caption text undergo OCR processing, while frames containing only non-caption text are excluded.
Solution Approach 2:
The trained neural network acts as an intermediary between the video frames and the OCR engine. It mediates the processing flow by filtering and selecting only the appropriate frames that contain burned-in captions, thereby improving the reliability of the overall text detection system and reducing false positives from non-caption text.
3Productivity
If a simple text recognition approach is used, then processing speed is fast, but accuracy in distinguishing burned-in caption text from non-caption text decreases
Solution Approach 1:
The processing system is segmented into two functional components: a fast neural network classifier for preliminary frame classification, and a slower but more accurate OCR engine for detailed text recognition. This segmentation allows the system to leverage the speed of the neural network for filtering while using the accuracy of OCR for final text identification, achieving both speed and precision.
Solution Approach 2:
The trained neural network serves as an intermediary that enhances the overall system accuracy. It pre-processes frames by identifying those with burned-in captions, thereby improving the input quality for the OCR engine and enabling more accurate discrimination between burned-in caption text and non-caption text while maintaining efficient processing speed.
Data Source
AI summary
In some embodiments, a method inputs a frame sample of a video into a prediction network of a discriminator. The frame sample is analyzed to determine whether the frame sample includes burned-in caption text. When the frame sample is determined to include burned-in caption text, the method sends the frame to a recognition engine to perform a recognition process on the frame sample, performs the recognition process on the frame sample to recognize text in the frame sample, and outputs the text for a service to be performed for the video. When the frame sample is determined to not include burned-in caption text, the method bypasses the recognition engine and does not perform the recognition process on the frame sample.