Image registration

In image registration, the pixels in an image are remapped to superimpose an arbitrary number of fixed control points. Image registration is thus a generalized form of affine transformation; however, in image registration, the parameters (translation, rotation, scaling, etc.) are not constant but vary with x and y.

Image registration is ideal for repairing gradual or ``differentiable'' distortions, but less suitable for correcting abrupt discontinuities in the image, such as those caused when an object moves while being scanned in a scanner. These can be removed by manual image warping (Sec. 7.3).

Registration of an image of sections of two protein 2D gels.
A,E. Original images with control data points labeled 1..6
B,F. Same images after obtaining data points. The points are marked with a small cross.
C,G. Vector display superimposed on image
D,H. Result of warping the right image. Program is in multiple cursor mode with cursor on spot #3. All spots in H are now aligned with their corresponding spots in D.

Typical uses for image registration would be in analyzing terrain maps, and 2D polyacrylamide gels. In a 2D gel, each protein spot is identified by its x and y coordinates, so it is necessary to know the exact position of each spot in order to accurately identify it. The figure above illustrates the steps in registering two protein 2D gel images.

The image registration algorithm in tnimage consists of two parts. First each part of the image is rotated to remove any rotations or twist in the image. Then, a large array of vectors or vector map is created that allows each part of the image to be elastically warped in the x,y direction to remove any residual differences in the pixel coordinates. This two-step procedure is used because rotation usually is much simpler than warping (i.e., one rarely finds images with regions of local twisting), and therefore rotation needs fewer control points. For example, if there is no local twisting, a single control point is sufficient to correct image rotation.

Once calculated, the same warping map can be used either to warp an image, or to locate the new positions of a set of data points. This approach is more flexible than calculating the new coordinates on an `as needed' basis because the same map can be used to warp multiple images very rapidly, to remap arbitrary lists of coordinates that may not be related to any image, and to create lists of coordinates for automatic densitometry without actually warping the image. This latter feature is important since warping an image can introduce slight variations in the size of features on an image.

Unlike other programs, tnimage does not need a predefined set of landmark points from the two images. Tnimage uses a powerful pattern-matching method that can usually find the correspondences between two sets of points without user intervention. Once the correspondences are found, the user can make corrections as needed before applying the map to an image.

Tnimage matches discrete control points instead of matching all grayscale features in an image. The reason for this is that grayscale features are irrelevant for many types of images (such as 2D gels). The control points can be obtained manually, or automatically by using the grain counting function in tnimage, or any other suitable means. For terrain maps, edges and corners can be used.

After the images are registered, the cursor can be set to ``multiple crosshairs'', which draws a set of crosshairs on each image. This facilitates identification of corresponding features. You can place the cursor at a spot in one image and observe the corresponding location of the second cursor on the other image. You can also run macros to perform densitometry (for example) on specified areas, without having to change the macro for each image.

The program uses two sets of data points:

• Data points - an arbitrarily large set of coordinates indicating the x and y locations of coordinates whose positions are to be identified. These can be read individually by clicking ``Read points'' or read from a combined file by clicking ``Read match table''.
• Landmark points - A smaller set of known points whose corresponding positions on both images are either already known or should be calculated. There is a practical maximum of 100 landmark points. The landmarks are stored in one single file.

The landmark points are used to calculate a vector map which can then be applied either to the data points or to an image. The landmark points do not need to be a subset of the data points.

Neither image has to be present unless you wish to obtain control points interactively or perform warping.

Clicking ``Correlate points'' automatically sorts and cross-correlates the two sets of landmark coordinates to determine the which landmark points correspond to each other, and eliminates orphan points. Correlating points is unnecessary if both landmark sets already contain corresponding points, as when the points have been obtained interactively.

Options

• Image no. - selects which image to use as a reference and as an unknown image. Only needed when obtaining points or warping images.
• Obtain landmarks - Manually obtain landmark points from reference and unknown image.
• Translate landmarks - Add an x and y value to the landmark points. This can be useful in improving the accuracy of the calculations.
• Read data points - reads a full set of data points to be remapped.
• Save data points - saves the data points after remapping.
• Restore data points - restores the original data points in case something went wrong with the remapping.
• Read landmarks - reads a set of landmark coordinates from disk.
• Edit landmarks - permits viewing and modification of the landmark coordinates.
• Save landmarks - save the modified landmarks list.
• Read match table - Read the worksheet containing the data points and other information from disk.
• Edit data/match table - Edit and modify the data points
• Save match table - Write the modified data points to disk.
• Correlate points - Calculate the correspondences between the possible landmark points and generate a landmark table (Not needed if the landmark points are already correlated).
• Calculate rotation map - Create a large rotation map suitable for rotating an image or remapping data points.
• Rotate image - rotate an image using the most recently calculated rotation map (which is derived from the landmark points).
• Smoothing - Sets the amount of smoothing to be applied to the vector map or rotation map (1=3, 2=5, 3=7, 4=9, 5=11, 6=15, 7=19, or 8=21 point smoothing). Changing this parameter can help in fine-tuning the construction of an optimal vector map.
• Calculate vector map - Create a large vector map suitable for warping an image or remapping data points.
• Warp image - warp an image using the most recently calculated warp vector map (which is derived from the landmark points).
• Show data vectors - Draws arrows indicating where each data point will move if the image is warped.
• Create unwarped spot list - Calculates the new identities of the current list of data points from the ``unknown'' data set along with their original coordinates, and displays them in an editor. This option should be selected if you did not warp the image.
• Create warped spot list - Calculates the new identities of the current list of data points from the ``unknown'' data set along with their new coordinates after warping, and displays them in an editor. This option should be selected if you warped the image.