UC Santa Cruz

Color-based computer vision approaches have proven pertinent in detecting and classifying objects in various areas ranging from industrial inspection to mobile vision and biomedical applications.

Smartphones are becoming an increasing research platform due to their high-quality cameras and programmability as well as their portability. The computational power of smartphones has been improving since the last decade which enables us to make use of them as a target detection and decoding device.

In this thesis, we use computer vision approaches to propose effective detection and decoding of multicolor surfaces. Our methods address the main issues related to designing a practical detection and decoding problems, namely robustness and computational efficiency.

To this end, this thesis offers contributions in multicolor detection and decoding in mobile vision.

In the first part of this thesis, we explore the potential use of smartphones to detect a special multicolor marker that could potentially help blind persons to find their way around in a suitably equipped environment. The use of multicolor surfaces not only increases the distinctiveness of the marker with respect to the background and thus more reliable detection, but also enables detection by a model-based method that explicitly takes into account the variability of illumination. In the second part of this thesis, we explore color information access by allowing users to decode a color barcode from a barcode image. Our image-based color barcode decoding approaches are motivated by the necessity of increasing information density in a limited space. In our approaches, we address practical issues such as changes of the observed colors due to changing illuminant, specular reflection, and blur-induced color mixing from neighboring barcode patches. In our initial decoding approaches, we consider groups of color surfaces that can be decoded under variety of illuminants, exploiting the fact that joint color changes can be represented by a low-dimensional subspace. Thus, decoding a group of color surfaces is equivalent to searching for the nearest subspace in a dataset. In another approach, rather than decoding individual patches or using a clustering method, our algorithm iteratively decodes the colors of all patches across the barcode image by minimizing the overall observation error. We achieve high information rates using only three reference colors with a very low probability of decoding error.