Feature extraction and automatic annotation of flash illuminated video data in unconstrained video streams

摘要

A method of processing video data (1), comprises the steps of extracting the video data (1) into a plurality of frames (2), and analysing the plurality of frames (2) for detecting a flash illumination in the video data (1). Therein, analysing the plurality of frames (2) comprises the steps of partitioning each frame of the plurality of frames (2) into a plurality of segments (31) of a predefined size and calculating an average segment luminance component (41) for each segment of the plurality of segments (31) of each frame of the plurality of frames (2). By partitioning the video frames (2) into smaller segments (31) and subsequently using the transition in luminance component in each segment, an efficient and fine-tuned detection of flash illumination in the video data (1) is possible. Thereby, even flash illumination not affecting the whole frame can be detected which can be of decisive importance. Thus, when applying the method according to the invention flash detection in unconstrained user generated video clips or scenes can be significantly improved and accordingly the video clip and its frames can efficiently and sophistically be classified or rated as containing flash-illumination or not. Also, when detecting frames (2) that contain scenes with flash illumination according to the invention, the risk of incorrect recognition of flash illuminated scenes, e.g., by recognising cut points as flash illuminations, of false alarm can be reduced.