QR Code Detection in Arbitrarily Acquired Images

Applications of Quick Response (QR) codes enable rich context interaction through creation of links between physical objects and Internet resources.

In spite of the widespread use of this kind of barcode, applications for visually impaired people and robots are not common because existing decoders assume that the symbol is properly framed during image acquisition.

This work proposes a two-stage component-based approach to perform accurate detection of QR code symbols in arbitrarily acquired images. In the first stage a cascade classifier to detect parts of the symbol is trained using the rapid object detection framework proposed by Viola-Jones. In the second stage, detected patterns are aggregated in order to evaluate if they are spatially arranged in a way that is geometrically consistent with the components of a QR code symbol. An extensive study of parameter variation of both stages was performed and the results were analyzed in terms of precision, recall and computational efficiency.

The proposed QR code detector achieved average recall of 91.7% with precision of 76.8% while being capable of processing a 640×480 pixels video stream at 22 fps. These results support implementation of real-time applications that assist visually impaired people and robots in mobile hardware, allowing them to have access to the wealth of information available through QR codes in multiple medium.