If a descriptor is not suitably robust when the size of the training set is small, the classifier may inappropriately separate the categories. For many applications this is infeasible (e.g., automated surveillance systems for detecting suspected criminals) and it could reduce execution speed for some non-parametric classification algorithms, e.g., nearest neighbor. One reason for this phenomenon is that an image recognition system can obtain invariance properties simply by training it with more data, as long as the additional data exhibits the same variation as the test set. Our experiments are conducted with relatively few instances, since classifier performance is much more sensitive to the quality of the descriptors in this setting. Our objective is to determine if the grayscale representation used significantly influences performance and if so, to identify which method is preferred regardless of the dataset or descriptor type. This comparison is performed using the Naive Bayes Nearest Neighbor (NBNN) image recognition framework and four different types of image descriptors. All of the methods are computationally inexpensive, i.e., they all have linear time complexity in the number of pixels. While we do not evaluate every method that has been developed, we evaluate all of the widely used methods, as well as some less well known techniques (e.g., Decolorize). In this paper we compare thirteen different methods for converting from color-to-grayscale. Indeed, color may be of limited benefit in many applications and introducing unnecessary information could increase the amount of training data required to achieve good performance. The main reason why grayscale representations are often used for extracting descriptors instead of operating on color images directly is that grayscale simplifies the algorithm and reduces computational requirements. A priori, none of these criteria suggest superior recognition performance.

The most common techniques are based on weighted means of the red, green, and blue image channels (e.g., Intensity and Luminance), but some methods adopt alternative strategies to generate a more perceptually accurate representation (e.g., Luma and Lightness ) or to preserve subjectively appealing color contrast information in grayscale images (e.g., Decolorize ).

However, since many methods for converting to grayscale have been employed in computer vision, we believe it is prudent to assess whether this assumption is warranted. This is because most researchers assume that the color-to-grayscale method is of little consequence when using robust descriptors. Modern descriptor-based image recognition systems often operate on grayscale images, with little being said of the mechanism used to convert from color-to-grayscale.

Gray scale image vs obj for cnc