91Èȱ¬

In order to classify the programmes we have developed several new ways of automatically analysing and classifying programmes. To begin, we take programmes that are available on iPlayer and use signal processing techniques on the audio and video of the programme to look for key characteristics of the programme. For example, we analyse the video for differences between images and the brightness of the video and also how loud different frequencies are in the audio. Once we have these features we can use them to help identify the mood of the programme. In 2011 we ran a series of user trials, asking people to watch clips of TV programmes and identify the change in mood as the clip progressed. We can then analyse this "ground truth" data to get a reference mood value and match analysed iPlayer programmes to mood values.

Whilst we’ve initially released this based on iPlayer content, our goal is that it will help people find new entertaining content from the 91Èȱ¬ Archives. Users can search for a programme that they know of and then find programmes from the archiveÌýwhich have similar (or opposite) moods. We hope to further extend this by adding topic similarity too – allowing users to find programmes with not only similar moods but also about similar things.

For detailed technical info on how we do this, I’ve listed some of our published papers at the bottom of this blog. If you have any queries please email us at

multimedia.classification@rd.bbc.co.uk

Ìý

Further reading;

• Affective Classification of Large Scale Broadcast Archives; 91Èȱ¬ R&D White Paper 201Ìý
• A Large Scale Experiment for Mood-based Classification of TV Programmes; 91Èȱ¬ R&D White Paper 232
• An Affective Interface for Mood-Based Navigation; 91Èȱ¬ R&D BlogÌý