LBP-inspired detection of color patterns: Multiplied local score patterns

Abstract

Local binary patterns (LBP) were previously used to characterize gray-scale patterns in an image. They have also been applied to color pattern recognition, but maintained a simple binary vector for classification. We have applied the sampling strategy of LBPs to collect local colors around every pixel. These samples are then individually scored with all models to find the best match. This determines the order the remaining color models are used to score the samples, leading to rotation invariance in a manner similar to LBPs. Once the scores are retrieved for each sample, they are modulated by the samples' saturation values. All modulated scores are then multiplied to produce a multiplied local score pattern (mLSP) map. Peaks are filtered based on their breadth using simple thresholding and subsequent connected component analysis. Results were gathered from 1534 images in two environments under two camera exposures, using two consumer printer technologies to produce the color pattern. The overall recognition rate was 86%. Recognition was further broken down to show effects of lighting environment, printer technology, camera distance, and color pattern setup. Pitfalls and potential solutions are discussed for the algorithm's use in a wider variety of environments and with other color patterns.