Logo Detection and Identification in System for Audio-Visual Broadcast Transcription

We present logo detection and identification based on a single-stage deep convolutional detector, the Scaled YOLOv4. This method is used in our system for audio-visual broadcast transcription and indexing which can be employed mainly for transcription of TV programs, mostly sports and advertising blocks. All transcribed information from audio and video streams together with time boundaries is indexed in the ElasticSearch database which can then be used to search for interesting keywords, entities etc. In this paper we present mainly the development and evaluation of the method for detection and identification of logos from images. We evaluate the logo detector on several of the most popular logo detection benchmarks, namely FlickrLogos-32, Logos-32plus, TopLogo-10 and QMUL-OpenLogo. The detector significantly outperforms the most common approach based on two stage models such as Faster R-CNN in terms of both speed and accuracy, achieving relative improvement up to 46% while running up to 2x faster.