Extracting data on product labels based on visual information and processing them in information systems opens the possibility of automating several logistic operations. We can pre-process the data for optical character recognition and barcode reading subsystems by localizing and segmenting the label in real-time.
In this thesis, I propose a solution, which segments the camera image by detecting interest points and fitting them on an existing model. The points can be reproduced at various affine transformations. They are selected based on features extracted based on the Hessian matrix of the intensity image and filtered by their chromaticity value. The interest points are then described by the SIFT descriptor, and matched against the model. After a successful segmentation the segmented parts of the image are handed over to the corresponding barcode reading of optical character recognition subsystem.
To explore the ability of using the method in real-time, I propose an implementation designed for embedded GPUs using OpenGL shaderprograms, as well as a set of algorithms to perform the registratoin, detection and decoding of the label. I present the details of formulating the algorithms and data representations for this combined GPU, CPU environment.
Measurements show that the proposed implementation runs fast enough to be used in real-time applications, and precise segmentation can be done using the selected interest points. The tests have been conducted on product labels of consumer electronics equipment.