DWT-based multi-scale clustering: Applications to scene-adaptive color quantization and auto white balancing

Data clustering is the process of grouping similar data samples into the same category. In the widely used clustering algorithms, the target number of cluster is assumed to be known. In many applications, however, the number of clusters is not defined and this number needs to be adjusted according to the nature of the data. To cope with this issue, Multi Scale Clustering (MSC) was developed by Kehtarnavaz et al. to determine the number of clusters automatically. However, the computational complexity associated with MSC avoids applying MSC to real-time applications. In this dissertation, the DWT (discrete wavelet transform)-based MSC is developed to address the computational complexity aspect of MSC and thus to allow its real-time deployment. To demonstrate the effectiveness and efficiency of the developed algorithm, it is applied to two color imaging applications: color quantization and auto white balancing (AWB). In the color quantization application, it is desired to reduce the number of colors in an image to a much smaller number of representative colors while keeping color distortion to an acceptable level. The reduction in the number of colors lowers computational complexity associated with color processing and achieves higher color image compression for storage and transmission purposes. The use of DWT-based MSC in a three-dimensional (3D) color space has lead to a scene-adaptive color quantization technique. The color quantization performance is evaluated in terms of compression ratio or number of representative colors, color distortion, and computational complexity. The results show that the developed technique outperforms the existing techniques in terms of color distortion while achieving comparable computational complexity. In the second application, the developed clustering algorithm is applied to the scoring based auto white balancing (AWB) for real-time deployment on digital still camera platforms. AWB involves the process of making white colors to appear as white under different illuminants. It is shown that the computational efficiency of DWT-based MSC along with several optimization steps allow the AWB processing to run in real-time on a digital camera processor. More specifically, the outcome of an actual implementation on the Texas Instruments TMS320DM320 processor is provided to illustrate the effectiveness of this approach in identifying the scene illuminant as compared to the widely used gray-world auto white balancing approach.