Color Invariance using Linear Image Filters

The problem with recognizing general shapes, such as balls, as in our particular case, is the number of possibilities in the visual appearance of a ball. A ball can take on any color and size and may have any pattern on its surface. In order to generalize the concept of a ball, the initial goal was to eliminate any color information in the data images representing the balls.

To detect the edges in the image, we use linear image filters followed by a threshold to eliminate noise data, which would then be given as input to the classifier, which in turn handles differences in size, pattern, lighting, etc. For this paper, we are using a Sobel filter, as described in [4].

In order to eliminate the color information in the images, we apply the filter to the colored image and then use a threshold $t$ to include any pixel in any of the 3 color channels that crossed the threshold $t$ value in the output image. The resulting image is a binary image including the thresholded pixels of the 3 color channels. A typical output image of this technique is shown in Fig. 3 (left).

This edge detection and thresholding technique is applied to all images used as input to the training of the Haar classifier. The training process is described in the following subsections.