Localization of Visual Codes in the DCT Domain Using Deep Rectifier Neural Networks

The reading process of visual codes consists of two steps, localization and data decoding. This paper presents a novel method for QR code localization using deep rectifier neural networks, trained directly in the JPEG DCT domain, thus making image decompression unnecessary. This approach is efficient with respect to both storage and computation cost, being convenient, since camera hardware can provide JPEG stream as their output in many cases. The structure of the neural networks, regularization, and training data parameters, like input vector length and compression level, are evaluated and discussed. The proposed approach is not exclusively for QR codes, but can be adapted to Data Matrix codes or other two-dimensional code types as well.