Updated all users of the binop_*.c files. Reworked pow to use the binop machinery.

Created special binary ops template for fpcomplex operations which are not element-wise (multiply, div), and updated the actual operators to use this. check-in: b0d870cef2 user: andreask tags: infinite-plane

unary: ahe, convolve, region_sum, rof

binary: atan2, hypot, pow, cannyinternal, complex div/multiply, join, joint-bilateral, threshold

other: warp-field, cut (*)

(*) Currently done via 'crop', semantic mismatch. Better as their own operator. check-in: 122bdc7d78 user: andreask tags: infinite-plane

Image notes

image, bbp/planes, meaning, results
---------------------------------------------------------------
lena   8/3 rgb full color - ok
lena2  8/1 rgb vga palette - ok - decoder now tested
lena3  4/1 16 color palette - not implemented
lena4  1/4 16 color palette, bits striped - not implemented
lena5  See lena2
lena6  1/3  8 color palette, bits striped - not implemented
lena7  1/2  4 color palette, bits striped - not implemented
lena8  2/1  4 color palette - not implemented
lena9  1/1  2 color black/white - ok, check for border issues.
lena10 See lena3
---------------------------------------------------------------

Added set of typed constructor taking location information. Updated the warp init code to make use of the new constructors.

Beginning (with float/float) to update the binop core codes to handle mismatched and -aligned images, i.e. of different sizes and translated to each other. Instead of using additional macros for the special cases we reuse the existing one and rely on the compiler to simplify this as much as possible (i.e. constants, irrelevant operations, etc.) check-in: f9955e8a0c user: andreask tags: infinite-plane

Embed images into the infinite 2d plane, i.e. make them bounded rectangles with a location in that plane. We had a start on that with the operators for projective warp, storing the origin point in the user-meta data. This information is now in the C level structures.

This commit provides the basic change, plus new accessors for the geometry data, and updates all users of said geometry data in C to use these new accessors.

The semantics of the various operations which could make use of the location information has not changed yet. This is for future commits in this branch. check-in: 1da2e834b8 user: andreask tags: infinite-plane

Having worked through and vetted most of the code in the parental branch I want this code in the main branches, for regular use.

But there is still unvetted code, namely the 'filter::cleanup' and the higher-order image registrations (rotation, scale, affine, find-object), where I have run into roadblocks of understanding myself, and where I defer to Kevin for vetting.

To get the vetted pieces into the mainline, without having the unvetted parts this commit is made. It removes the unvetted pieces, puts them into a new branch, and this new branch will then be integrated with "works-for-critcl-2.1" and then "works-for-critcl-3".

The code is not lost, of course. It is still in the "x-smsaurabh" branch, and when Kevin has worked through it, it can then be integrated also. Leaf check-in: d7015891a3 user: andreask tags: integration-stage

Fixed the gausian noise as well, which has to use two uniformly distributed random numbers (radius, angle) to get one gaussian distributed value. check-in: 4b1c8ccdd5 user: andreask tags: x-smsaurabh

Moved the complex <-> float conversion to the 'convert' ensemble.

Extended the FOO -> complex conversion to directly allow greyN images as input.

I.e. we can now do a 1-step greyN -> complex conversion, instead of requiring the 2-step greyN -> float -> complex

Dropped the fpcomplex2real primitive, this is the same as 'convert 2float', and now implemented using the primitive for the conversion.

Extended 'split' and 'join' for image type 'fpcomplex'.

Updated all users, updated the documentation. check-in: bce7e722de user: andreask tags: x-smsaurabh

Swapped this operator's arguments as well. When width and height are used together they are pretty much everywhere used in this order (w, h). (Like in the result of crimp::dimensions). This enabled the simplification of the calls (single 'dimensions' instead of separate 'width' and 'height').

Exposed the operator as 'crimp noise random'; updated the documentation, and added demos. check-in: 08bc13765e user: andreask tags: x-smsaurabh

Plus fixes in the BMP reader: Bad return value, and data alignment when decoding pixel data is relative to the base location. Error messages modified to uniquely identify each location and problem. check-in: fc7e925cfb user: andreask tags: image-io-experiments

This replaces rev [1a2b5e973a] (moved to a mistake branch), and reverts part of rev [2e4a6fce26].

This change relies on functionality provided by critcl in rev [8299af2478] and forward. check-in: 62fc511af8 user: andreask tags: work-for-critcl-2.1

Furthermore fixed the issue of the missing M_PI definition as well. check-in: f3474ad7e1 user: andreask tags: work-for-critcl-2.1

Thanks to Pat Thoyts for the report. check-in: 7a934fa654 user: andreask tags: trunk

Modified the symmetric binary operators to not try argument reversal when the argument image types are identical, as then the second try is identical to the first, thus superfluous.

Filled the conversion of gradients from cartesian to polar representation, and added a method to generate a visualization of the gradient. check-in: 1d969fb0c3 user: andreask tags: ak-experimental

This was done because usage of the PGM format for the greyscale page images in the bookflow project showed that the conversion of the read PGM data to a crimp image was a big bottleneck, dominating the time to calculate the global brightness of the pages.

Using one the images as example (IMG_2291.pgm, 3264x2448) let us compare the before and after times for all the steps, for both the raw and plain sub-formats of PGM:

Much better. The time spent on calculating the statistics and reading the bytes is essentially the same, within error margins, and the parsing of the PGM format is an order of magnitude faster. check-in: 186b7b5109 user: andreask tags: ak-experimental

Applications of these are whitepoint correction, conversion between various colorspaces (CIE *, LMS, ...), etc.

Added convenience commands for various colorspace conversion matrices (RGB -> CIE XYZ, CIE XYZ -> LMS, RGB -> LMS (analogous to previous)).

Still need the inverse conversions.

Started on the CIE spaces, this is very incomplete, and the value ranges (domains) especially are not known, so the scaling is very likely very off.

Plus demos. check-in: 1cc2e4740f user: andreask tags: ak-experimental

Added the extended statistics to the hough demo. check-in: fc24897d1e user: andreask tags: ak-experimental

• [60254f4954] arbitrary affine transformations.
• [7068afa5f5] rotation through arbitrary angles and center.
• [f17efe0938] arbitrary projective transformations.
• [233e0d0b5a] general deformation through quadrilateral mapping.
• [126d0ace8f] bilinear/bicubic interpolation.

Re-added a resize method, now based on projective warp.

Renamed the 'crimp warp warp' to 'crimp warp field', in deference to the vector fields providing the sampling coordinates. And yes, it is a bit tongue-in-cheek re StarTrek.

Implemented the primitives for 'warp field'.

Demo for 'warp field', a blur mode like light scattering through milky acryl.

Documented 'resize, 'warp field' and the 'transform' methods.

New demos for quadrilateral mapping. check-in: 5a58eae505 user: andreask tags: warp-experiment

Even with the gcd to reduce sample factors it is to easy to run into large upsampling factors and massive memory consumption by the transient intermediate images. Just need new/old width/height which are relative prime (gcd == 1).

We need something which has all the three phases (upsample, convolve, downsample) integrated, avoiding intermediate full images.

Starting to look towards the 'matrix' primitive and command now. More general (projective transform), and having the resampling integrated into the output loop. It needs some fixing tough, as is it is not usable for the intended purpose.

When that is working however we can make resize and related commands simple wrappers around it, for convenience. check-in: bbe91135d7 user: andreask tags: ak-experimental

The implementation uses summed area tables, i.e. integral images, to calculate the sum of pixels in all windows, making it very fast as the summation is independent of the filter radius, i.e. window size. check-in: cb6683b4d3 user: andreask tags: ak-experimental

This completes the implementation of ticket [723f624229854bf48c120a238ffc23f3cf4f96eb]. check-in: 5f4c3be4ba user: andreask tags: ak-experimental

1. Reworked the statistics method to be an ensemble. The original statistics command now is 'statistics basic'. Extended this command to compute the mid-grey (middle) value (1/2(min+max)).
2. Changed the dictionary a bit, use a single 'channel' sub-dictionary to hold the per-channel statistics, instead of a flat hierarchy, with the hierarchy in long keys. Updated the statistics demo to take this into account for printing. Ditto the threshold-le demo.
3. Added an otsu subcommand to compute the binarization threshold of each channel in a statistics dictionary via Otsu's method.
4. Added global thresholding commands which compute the threshold to use on their own, in various ways (mean/average, median, middle, otsu).
5. Added a separate demo for the global thresholding commands with automatic threshold.
6. Updated the documentation for the statistics and global thresholding commands.

Still have to go through the primitives and decide how each handles this meta data from input to output.

Still have to implement some nice handling of the data on top of the meta primitives (i.e. treat as a dictionary). check-in: 48ba899215 user: andreask tags: ak-experimental

(1) The primitives now expose the expected image-type of their kernel through their name.

(1a) Updated the 'filter convolve' implementation to select the primitive based on kernel type, and throw errors if no primitives are available for image and kernel types.

(2) Extended the existing set of primitives (using a grey8 signed kernel) to grey16, grey32, and float image types.

(3) Created a whole new set of primitives which use kernels specified in the new floating-point image-type.

(4) Extended the kernel ensemble with commands to make FP-kernels.

(5) New demo, family of gaussian kernels, sampled and discrete. check-in: 121b03874c user: andreask tags: ak-experimental

Created the whole suite of flip primitives for it. This implicitly provides us with the quarter and half rotations as well. check-in: 17013beeee user: andreask tags: ak-experimental

(1) Modified the read-tcl primitive to expose the destination type in its name.

(2a) New read-tcl primitive with the new float type as destination format.

(3) Updated the 'read' API initialization to handle the expansion of destination types for 'read tcl'.

(4) Updated all users of 'read tcl' to be 'read tcl grey8', code and demos, i.e. properly announce the destination type. check-in: 06979e6ed5 user: andreask tags: ak-experimental

Fixed problems in demo shutdown exposed by that (writing of demo times, and the defered compile+loading of C-level primitives).

Modified the demo startup to begin with no images selected. check-in: bf3c48bcb1 user: andreask tags: ak-experimental

Extended the demos with laplace operators, embossing, and basic edge detectors (simple and sobel). check-in: 00eb759f1f user: andreask tags: ak-experimental

Fixed ticket [7dea5be1bf9f099bfa46614ba5992c51245a8805]. check-in: 882ebaeb7a user: andreask tags: ak-experimental

Pulled the helper code for cumulation and scaling from the equalization demos. Found and fixed a bug in the cumulation. check-in: 3730fe3d94 user: andreask tags: ak-experimental

Updated demos and documentation. Plus more demos for rank filtering (min, and max). check-in: a526962f88 user: andreask tags: ak-experimental

Updated demos and documentation. check-in: 59d925e147 user: andreask tags: ak-experimental

Added the RGB and HSV variants. check-in: e2992a1a13 user: andreask tags: ak-experimental

Simplified the primitive, ripped out the special case code handling of image borders. This means that to avoid accessing pixels outside of the input image we scan only the inner pixels of the input, and consequently get a result shrunken by the kernel dimensions minus one.

To compensate for that we use the expand primitives implemented in the last revisions to make the input appropriately larger before applying the convolution.

The convolve method now accepts an option -border enabling the caller to specify what type of border to use, with the default zero const border. This shares the border type processing with 'expand', therefore the code for that was moved into its own internal helper command.

The method was further extended to accept more than one kernel, and apply them in sequence.

The methods to make and transpose kernels have been reorganized into an ensemble. check-in: 3cc7886413 user: andreask tags: ak-experimental

Added keywords to the documentation of various methods.

Moved the figures into their directory, and the math figures again in a subdirectory of that. check-in: f84c9769cf user: andreask tags: ak-experimental

Reworked the handling of border-type specific arguments for const: auto-extend missing data per number of color channels in the image), ignore superfluous values.

Renamed 'extend' to 'replicate', so that we can use 'extend' for the last border type to be implemented. check-in: cdbfcd8700 user: andreask tags: ak-experimental

* Replication of the border pixels.

Not yet done are:

* Subtract reflection from the border pixel. check-in: 80aa718ff4 user: andreask tags: ak-experimental

* Adding a constant pixel. * Cyclically (toroidal) wrap around the borders. * Reflection/mirroring at the borders.

Not yet done are:

* Replication of the border pixels. * Subtract reflection from the border pixel. check-in: d76e4838c1 user: andreask tags: ak-experimental

Added template code for all the missing directions between rgb, rgba, and grey8.

Added the missing operator implementations. For the commutative operators we are leaving out half of the missing, and handle them through commutation in the Tcl level. check-in: 2c2c0c2211 user: andreask tags: ak-experimental

The demos are now handled like they are in a matrix, with the choice of images and actions as the two axes.

Choose an action and it applies to the current image. Choose an image and the current action is applied.

So, choose an image and walk the toolbar to see the various actions. Or, choose an action, maybe parameterize it, and then walk the images. check-in: 8ee78df8f1 user: andreask tags: ak-experimental