Effects of Classification Methods on Color-Based Feature Detection With Food Processing Applications

Abstract

Color information is useful in vision-based feature detection, particularly for food processing applications where color variability often renders grayscale-based machine-vision algorithms that are difficult or impossible to work with. This paper presents a color machine vision algorithm that consists of two components. The first creates an artificial color contrast as a prefilter that aims at highlighting the target while suppressing its surroundings. The second, referred to here as the statistically based fast bounded box (SFBB), utilizes the principal component analysis technique to characterize target features in color space from a set of training data so that the color classification can be performed accurately and efficiently. We evaluate the algorithm in the context of food processing applications and examine the effects of the color characterization on computational efficiency by comparing the proposed solution against two commonly used color classification algorithms; a neural-network classifier and the support vector machine. Comparison among the three methods demonstrates that statistically based fast bounded box is relatively easy to train, efficient, and effective since with sufficient training data, it does not require any additional optimization steps; these advantages make SFBB an ideal candidate for high-speed automation involving live and/or natural objects. Note to Practitioners-Variability in natural objects is usually several orders of magnitude higher than that for manufactured goods and has remained a challenge. As a result, most solutions to inspection problems of natural products today still have humans in the loop. One of the factors influencing the success rate of color machine vision in detecting a target is its ability to characterize colors. When unrelated features are very close to the target in the color space, which may not pose a significant problem to an experienced operator, they appear as noise and often result in false detection. This paper illustrates the applicability of the algorithm with a number of representative automation problems in the context of food processing applications. As demonstrated experimentally, the artificial color contrast and statistically based fast bounded box methods can significantly improve the success rate of the detection by reducing the standard deviation of both the target and noise pixels, enlarging the separation between feature clusters in color space, and more tightly characterize the feature color from its background. The algorithm presented here has several advantages, including simplicity in training and fast classification, since only three simple checks of rectangular bounds are performed