create_shape_model — Prepare a shape model for matching.
create_shape_model(Template : : NumLevels, AngleStart, AngleExtent, AngleStep, Optimization, Metric, Contrast, MinContrast : ModelID)
The operator create_shape_model prepares a template, which
is passed in the image Template, as a shape model used for
matching. The ROI of the model is passed as the domain of
Template.
The output parameter ModelID is a handle
for this model, which is used in subsequent calls to
find_shape_model. The center of gravity of the domain (region) of the
model image Template is used as the origin (reference point) of the
model. A different origin can be set with set_shape_model_origin.
The model is generated using multiple image pyramid levels and is
stored in memory. If a complete pregeneration of the model is
selected (see below), the model is generated at multiple rotations
on each level. The model can be extended by clutter parameters with
set_shape_model_clutter.
The number of pyramid levels is determined with the parameter
NumLevels. It should be chosen as large as possible
because by this the time necessary to find the object is
significantly reduced. On the other hand, NumLevels must
be chosen such that the model is still recognizable and contains a
sufficient number of points (at least four) on the highest pyramid
level. This can be checked using the output of
inspect_shape_model. If not enough model points are
generated, the number of pyramid levels is reduced internally until
enough model points are found on the highest pyramid level. If this
procedure would lead to a model with no pyramid levels, i.e., if the
number of model points is already too small on the lowest pyramid
level, create_shape_model returns with an error message.
If NumLevels is set to 'auto' (or 0 for
backwards compatibility), create_shape_model determines the
number of pyramid levels automatically. The automatically computed
number of pyramid levels can be queried using
get_shape_model_params. In rare cases, it might happen that
create_shape_model determines a value for the number of
pyramid levels that is too large or too small. If the number of
pyramid levels is chosen too large, the model may not be recognized
in the image or it may be necessary to select very low parameters
for MinScore or Greediness in find_shape_model
in order to find the model. If the number of pyramid levels is chosen too
small, the time required to find the model in
find_shape_model may increase. In these cases, the number
of pyramid levels should be selected using the output of
inspect_shape_model.
The parameters AngleStart and AngleExtent
determine the range of possible rotations, in which the model can
occur in the image. Note that the model can only be found in this
range of angles by find_shape_model. The parameter
AngleStep determines the step length within the selected
range of angles. Hence, if subpixel accuracy is not specified in
find_shape_model, this parameter specifies the accuracy that
is achievable for the angles in find_shape_model.
AngleStep should be chosen based on the size of the object.
Smaller models do not possess many different discrete rotations in
the image, and hence AngleStep should be chosen larger for
smaller models. If AngleExtent is not an integer multiple
of AngleStep, AngleStep is modified accordingly.
To ensure that for model instances without rotation angle values of
exactly 0.0 are returned by find_shape_model,
the range of possible rotations is modified as follows:
If there is no positive integer
value n such that AngleStart plus n times
AngleStep is exactly 0.0, AngleStart is decreased
by up to AngleStep and AngleExtent is increased by
AngleStep.
For particularly large models, it may be useful to reduce the number
of model points by setting Optimization to a value
different from 'none'. If Optimization =
'none', all model points are stored. In all other cases,
the number of points is reduced according to the value of
Optimization. If the number of points is reduced, it may
be necessary in find_shape_model to set the parameter
Greediness to a smaller value, e.g., 0.7 or 0.8. For small
models, the reduction of the number of model points does not result
in a speed-up of the search because in this case usually
significantly more potential instances of the model must be
examined.
If Optimization is set to 'auto',
create_shape_model automatically determines the reduction of
the number of model points.
The parameter Metric determines the conditions under which
the model is recognized in the image.
If Metric =
'use_polarity', the object in the image and the model must
have the same contrast. If, for example, the model is a bright
object on a dark background, the object is found only if it is also
brighter than the background.
If Metric =
'ignore_global_polarity', the object is found in the image
also if the contrast reverses globally. In the above example, the
object hence is also found if it is darker than the background. The
runtime of find_shape_model will increase slightly in this
case.
If Metric = 'ignore_local_polarity', the
model is found even if the contrast changes locally. This mode can,
for example, be useful if the object consists of a part with medium
gray value, within which either darker or brighter sub-objects lie.
Since in this case the runtime of find_shape_model increases
significantly, it is usually better to create several models that
reflect the possible contrast variations of the object with
create_shape_model, and to match them simultaneously with
find_shape_models.
The above three metrics can only be applied to single-channel images. If a multichannel image is used as the model image or as the search image, only the first channel will be used (and no error message will be returned).
If Metric = 'ignore_color_polarity', the model is
found even if the color contrast changes locally. This is, for
example, the case if parts of the object can change their color,
e.g., from red to green. In particular, this mode is useful if it
is not known in advance in which channels the object is visible. In
this mode, the runtime of find_shape_model can also increase
significantly. The metric 'ignore_color_polarity' can be
used for images with an arbitrary number of channels. If it is used
for single-channel images it has the same effect as
'ignore_local_polarity'. It should be noted that for
Metric = 'ignore_color_polarity' the number of
channels in the model creation with create_shape_model and
in the search with find_shape_model can be different. This
can, for example, be used to create a model from a synthetically
generated single-channel image. Furthermore, it should be noted
that the channels do not need to contain a spectral subdivision of
the light (like in an RGB image). The channels can, for example,
also contain images of the same object that were obtained by
illuminating the object from different directions.
The parameter Contrast determines the contrast the model
points must have. The contrast is a measure for local gray value
differences between the object and the background and between
different parts of the object. Contrast should be chosen
such that only the significant features of the template are used for
the model. Contrast can also contain a tuple with two
values. In this case, the model is segmented using a method similar
to the hysteresis threshold method used in edges_image.
Here, the first element of the tuple determines the lower threshold,
while the second element determines the upper threshold. For more
information about the hysteresis threshold method, see
hysteresis_threshold. Optionally, Contrast can
contain a third value as the last element of the tuple. This value
determines a threshold for the selection of significant model
components based on the size of the components, i.e., components
that have fewer points than the minimum size thus specified are
suppressed. As the minimum size is applied on the extent of the components,
the derived model contours can still be smaller than the specified minimum
size. This threshold for the minimum size is divided by two
for each successive pyramid level. If small model components should
be suppressed, but hysteresis thresholding should not be performed,
nevertheless three values must be specified in Contrast.
In this case, the first two values can simply be set to identical
values. The effect of this parameter can be checked in advance with
inspect_shape_model.
If Contrast is set to 'auto', create_shape_model
determines the three above described values automatically.
only the contrast ('auto_contrast'), the hysteresis thresholds
('auto_contrast_hyst'), or the minimum size
('auto_min_size') can be determined automatically. The
remaining values that are not determined automatically can
additionally be passed in the form of a tuple. Also various
combinations are allowed: If, for example,
['auto_contrast','auto_min_size'] is passed, both the
contrast and the minimum size are determined automatically. If
['auto_min_size',20,30] is passed, the minimum size is
determined automatically while the hysteresis thresholds are set to
20 and 30, etc. In certain cases, it might happen
that the automatic determination of the contrast thresholds is not
satisfying. For example, a manual setting of these parameters should
be preferred if certain model components should be included or
suppressed because of application-specific reasons or if the object
contains several different contrasts. Therefore, the contrast
thresholds should be automatically determined with
determine_shape_model_params and subsequently verified using
inspect_shape_model before calling
create_shape_model. Note that MinContrast influences the
automatic contrast estimation, and hence also the estimation of the minimum
size.
With MinContrast, it can be determined which contrast the
model must at least have in the recognition performed by
find_shape_model. In other words, this parameter separates
the model from the noise in the image. Therefore, a good choice is
the range of gray value changes caused by the noise in the image.
If, for example, the gray values fluctuate within a range of 10 gray
levels, MinContrast should be set to 10. If multichannel
images are used for the model and the search images, and if the
parameter Metric is set to 'ignore_color_polarity'
(see above) the noise in one channel must be multiplied by the
square root of the number of channels to determine
MinContrast. If, for example, the gray values fluctuate
within a range of 10 gray levels in a single channel and the image
is a three-channel image MinContrast should be set to 17.
Obviously, MinContrast must be smaller than
Contrast. If the model should be recognized in very low
contrast images, MinContrast must be set to a
correspondingly small value. If the model should be recognized even
if it is severely occluded, MinContrast should be slightly
larger than the range of gray value fluctuations created by noise in
order to ensure that the position and rotation of the model are
extracted robustly and accurately by find_shape_model.
If MinContrast is set to 'auto', the minimum contrast
is determined automatically based on the noise in the model
image. Consequently, an automatic determination only makes sense if
the image noise during the recognition is similar to the noise in
the model image. Furthermore, in some cases it is advisable to
increase the automatically determined value in order to increase the
robustness against occlusions (see above). The automatically
computed minimum contrast can be queried using
get_shape_model_params.
Optionally, a second value can be passed in Optimization.
This value determines whether the model is pregenerated completely
or not. To do so, the second value of Optimization must be
set to either 'pregeneration' or
'no_pregeneration'. If the second value is not used (i.e.,
if only one value is passed), the mode that is set with
set_system('pregenerate_shape_models',...) is used. With
the default value ('pregenerate_shape_models' =
'false'), the model is not pregenerated completely. The
complete pregeneration of the model normally leads to slightly lower
runtimes because the model does not need to be transformed at
runtime. However, in this case, the memory requirements and the
time required to create the model are significantly higher. It
should also be noted that it cannot be expected that the two modes
return exactly identical results because transforming the model at
runtime necessarily leads to different internal data for the
transformed models than pregenerating the transformed models. For
example, if the model is not pregenerated completely,
find_shape_model typically returns slightly lower scores,
which may require setting a slightly lower value for MinScore than
for a completely pregenerated model. Furthermore, the poses
obtained by interpolation may differ slightly in the two modes. If
maximum accuracy is desired, the pose of the model should be
determined by least-squares adjustment.
If a complete pregeneration of the model is selected,
the model is pregenerated for the selected angle range and stored
in memory. The memory required to store the model is proportional
to the number of angle steps and the number of points in the model.
Hence, if AngleStep is too small or AngleExtent
too big, it may happen that the model no longer fits into the
(virtual) memory. In this case, either AngleStep must be
enlarged or AngleExtent must be reduced. In any case, it
is desirable that the model completely fits into the main memory,
because this avoids paging by the operating system, and hence the
time to find the object will be much smaller. Since angles can be
determined with subpixel resolution by find_shape_model,
AngleStep >= 1 can be
selected for models of a diameter smaller than about 200 pixels.
If AngleStep =
'auto' (or 0 for backwards compatibility) is
selected, create_shape_model automatically determines a
suitable angle step length based on the size of the model. The
automatically computed angle step length can be queried using
get_shape_model_params.
If a complete pregeneration of the model is not selected, the model
is only created in a reference pose on each pyramid level. In this
case, the model must be transformed to the different angles and
scales at runtime in find_shape_model. Because of this, the
recognition of the model might require slightly more time.
Note that pregenerated shape models are tailored to a specific image size. For runtime reasons using images of different sizes during the search with the same model in parallel is not supported. In this case, copies of the same model must be used, otherwise the program may crash!
This operator returns a handle. Note that the state of an instance of this handle type may be changed by specific operators even though the handle is used as an input parameter by those operators.
Template (input_object) (multichannel-)image → object (byte / uint2)
Input image whose domain will be used to create the model.
NumLevels (input_control) integer → (integer / string)
Maximum number of pyramid levels.
Default value: 'auto'
List of values: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 'auto'
AngleStart (input_control) angle.rad → (real)
Smallest rotation of the pattern.
Default value: -0.39
Suggested values: -3.14, -1.57, -0.79, -0.39, -0.20, 0.0
AngleExtent (input_control) angle.rad → (real)
Extent of the rotation angles.
Default value: 0.79
Suggested values: 6.29, 3.14, 1.57, 0.79, 0.39
Restriction: AngleExtent >= 0
AngleStep (input_control) angle.rad → (real / string)
Step length of the angles (resolution).
Default value: 'auto'
Suggested values: 'auto', 0.0175, 0.0349, 0.0524, 0.0698, 0.0873
Restriction: AngleStep >= 0 && AngleStep <= pi / 16
Optimization (input_control) string(-array) → (string)
Kind of optimization and optionally method used for generating the model.
Default value: 'auto'
List of values: 'auto', 'no_pregeneration', 'none', 'point_reduction_high', 'point_reduction_low', 'point_reduction_medium', 'pregeneration'
Metric (input_control) string → (string)
Match metric.
Default value: 'use_polarity'
List of values: 'ignore_color_polarity', 'ignore_global_polarity', 'ignore_local_polarity', 'use_polarity'
Contrast (input_control) number(-array) → (integer / string)
Threshold or hysteresis thresholds for the contrast of the object in the template image and optionally minimum size of the object parts.
Default value: 'auto'
Suggested values: 'auto', 'auto_contrast', 'auto_contrast_hyst', 'auto_min_size', 10, 20, 30, 40, 60, 80, 100, 120, 140, 160
MinContrast (input_control) number → (integer / string)
Minimum contrast of the objects in the search images.
Default value: 'auto'
Suggested values: 'auto', 1, 2, 3, 5, 7, 10, 20, 30, 40
Restriction: MinContrast < Contrast
ModelID (output_control) shape_model → (handle)
Handle of the model.
If the parameters are valid, the operator create_shape_model
returns the value 2 (H_MSG_TRUE). If necessary an exception is raised. If
the parameters NumLevels and Contrast are chosen
such that the model contains too few points, the error 8510 is
raised.
draw_region,
reduce_domain,
threshold
find_shape_model,
find_shape_models,
get_shape_model_params,
clear_shape_model,
write_shape_model,
set_shape_model_origin,
set_shape_model_clutter
create_scaled_shape_model,
create_aniso_shape_model
Matching