The Hough transform is used to identify Kikuchi bands from the diffraction pattern. It is an image processing algorithm to facilitate the detection of lines inside an image. The transformation converts a line (1 pixel thick) in the image space into a point in the Hough space.
The Hough transform operation takes two parameters, the two resolutions of the Hough transform ( and ).
The image space is the one of the diffraction pattern where the origin is taken at the center of the image. It is a discrete space made up of a certain amount of pixels in the x and y directions. The intensity of those pixels can be seen as the third dimension. Similarly, the Hough space has three dimensions. The x and y axes of the image space are replaced by the and axes while the third dimension now represents the intensity of the Hough space. By definition, the Hough space is continuous since within their boundaries and can take any value. The Hough space is quantized to allow for computerized treatment. As for the image space, the discrete Hough space has a certain amount of pixels in the and direction, namely and .
The transformation is performed by calculating using the following equation for each pixel in the image space and for each in the Hough space, where the subscript i refers to the index of the pixels in the image space and j to the index of the pixels in the Hough space [1].
Effectively, this transformation converts each pixel of the image space into a sinusoidal curve in the Hough space. The calculated value is rounded to the closest pixel . The intensity of the pixels that are part of the sinusoidal curve are augmented by the intensity of the corresponding pixel in the image space. The accumulation of these intensities give rise to peaks in the Hough space which corresponds to the and coordinates of the bands in the image space.
The understanding of these results is not straightforward. An obvious question is why sinusoidal curves of individual, uncorrelated pixels in a band intersect in the Hough space at a specific and unique position? To answer this question, we shall refer to the following figure where the image and Hough space are respectively shown on the left and right of the figure.
From the definition of the Hough transform, each pixel in the image space is transformed into a sinusoidal curve in the Hough space. The curve represents all the possible unidimensional lines that can be passing through that pixel in the image space. A few lines are drawn in the figure above with their corresponding position in Hough space represented by circle markers. Only a small fraction of the lines are fully contained in the band, the rest of the lines cross it, but most of their pixels are outside the band.
If this geometrical construction is repeated for another pixel, B, of the band L, the same result is obtained. In the following figure, the lines passing by B and their equivalent representation in Hough space using triangular marker. All the lines or curves related to pixel B are drawn as dashed lines.
The lines inside of band L and passing by pixel B are the same lines that are also passing by pixel A. In Hough space, these lines end up having the same coordinates and , forming a peak. The intersection of the sinusoidal curves therefore corresponds to the lines that are fully inscribed inside the band in the image space. The intensity at this intersection is higher than the background because of two interlinked reasons:
- the sinusoidal curve of the pixels in the band have a higher intensity that the one of the pixels outside of it
- the intensity of many sinusoidal curves is added at this intersection.
If the band would have a width of 1 px, the area covered by its corresponding peak in Hough space would be approximately equal to 1 px:sup:2 [2]. However, the bands in a diffraction pattern are wider than 1 px. This results in the formation of a peak covering a large area. The center of a peak corresponds to the center of its corresponding band. From our previous explanation, the height and width of the peak will depend on the lines that pass through the pixels of the band and that are fully inscribed inside it. The operation Auto Hough transform tries to minimize this phenomenon by properly selecting the for a given . More explanations are given in the operation page.
Moving away from the conceptual Hough transform, the following figures show an experimental diffraction pattern of a silicon single crystal and its Hough space representation.
The location of the most intense Kikuchi bands can be clearly identified in Hough space by the bright peaks while other peaks are more faint and barely noticeable. It is the task of the peak detection algorithm to segment out the high intensity peaks from the background and disregard possible false peaks. The segmentation of the Hough space is shown in the following figure:
To evaluate the result, the corresponding line of each peak in the previous figure is overlaid on the original diffraction pattern. The lines and the peaks are colour-coded to illustrate their relationship.
In EBSD-Image, is varied between and can take value between . The width and height of the HoughMap are adjusted according to these ranges.
To prevent biasing effects as reported by Tao & Eades (2005) [3], the intensity at each coordinate and in the Hough space is equal to the average (instead of the sum as originally described by Krieger Lassen [2]) intensity of all the sinusoidal functions passing through this coordinate. The intensity of a coordinate in the Hough space is therefore the average intensity of the pixels along its corresponding line in the image space.