List of Operators ↓
This chapter describes the operators for texture inspection.
The texture inspection model enables the inspection of textured surfaces with only having to set few parameters. The algorithm requires images of flawless texture for training. The training process extracts texture features from the training images and trains a texture inspection model that is based on Gaussian Mixture Model (GMM) classifiers. After a successful training, texture images can be compared to the texture inspection model and potential defects can be identified. Texture inspection works on image pyramids. This way, multiple frequency ranges of the texture are analyzed.
It is possible to use multi-channel images in the texture inspection pipeline.
In the following, the steps that are required to perform the texture inspection are described briefly.
In a first step, a texture inspection model is created with
or read with
adds sample images to the texture inspection model.
Images that have been added to the texture inspection model with the help
can be viewed with
. If all or a part of the added
images are not needed for the current texture inspection, they can be removed
from the model using
The texture inspection model is trained with
In the training process an image pyramid is created. For each pyramid
level a Gaussian Mixture Model
(GMM) is trained and a
'novelty_threshold' is determined. The
'novelty_threshold' helps to distinguish between flawless and
defective texture. After the training, it is possible to query the
novelty thresholds with
Several parameters influence the training. They can be set with
We recommend the following proceeding for the training:
Set initial parameters for the training, if needed:
'patch_normalization': Setting this parameter to
'weber', the features of the used texture patches are
normalized making the texture inspection model more robust against
illumination changes. Note, however, that sometimes a normalization
might not be desired, e.g., if the brightness is needed to
discriminate between flawless and defective parts. In that case, a
controlled illumination is absolutely necessary for a successful
'patch_rotational_robustness': Setting this parameter to
'true', the features of the used texture patches are sorted in
a way that is robust against rotational changes. This can be used if
it is difficult to train all possible orientations of the textured
surface via training samples.
Train the texture inspection model (using the default parameters): The initial training of the texture inspection model may be very time consuming, depending on the number of training images and their resolution. The following time saving strategies can be considered:
Zoom the images: Often, the texture inspection works even better on low resolution images. If it is known that the full resolution of the images is not needed, you can significantly reduce the training time by zooming the images. Zooming the image down by a factor of 2, e.g., reduces the training time by factor 4.
Accelerate the training: Another, but more risky strategy is to
increase the threshold for the stopping criterion of the classifier
optimization by setting the parameter
e.g., from 0.001 to 0.1. This leads to shorter
training times but also to a possibly less accurate texture inspection
model. If this strategy is used, we strongly recommend to reduce
'gmm_em_threshold' again after most of the parameters have been
Check the result of the training: Use a set of images with defective textures, classify them with the trained texture inspection model and check the results as is described for the step 'Apply the texture inspection model'. If the results are not satisfying, further parameters can be adjusted for the training. The most important parameters are:
'levels': The number of used
pyramid levels or the explicitly used pyramid levels. Higher pyramid
levels work on coarser texture features. If the texture in your images
is very coarse, the lower pyramid levels may not be needed for a
successful inspection. Then runtime can be reduced by setting the
pyramid levels of interest explicitly with
'levels'. If for
'levels' are set to [1,3], the training and the
texture classification are only performed on the first and third level.
'sensitivity': Controls how strict the novelty thresholds
are set. This way, the sensitivity to novelties can be adjusted. In
general, positive values lead to fewer detected defects and negative
values lead to less detected defects.
'novelty_threshold': Sets the novelty threshold manually.
This can be used for fine tuning if the results of the automatic
estimation are not optimal.
Repeat the training, if needed: If parameters were adjusted, usually the texture inspection model has to be retrained.
After successfully training the texture inspection model it can be used to classify textured surfaces. Each test image can be compared to the texture inspection model with
For each test image the operator returns a
indicates where the image has high deviations to the trained texture samples.
'novelty_region' is generated by
combining the novelty regions of the different pyramid levels. The novelty
regions of adjacent levels in the image pyramid are intersected with each
other. This step helps to improve the robustness towards noise within the
single pyramid level responses. The intersected novelty regions are then
added to the returned
'novelty_region'. If a pyramid level has no
adjacent pyramid level, the region itself is added to the
'novelty_region'. If, for example,
'num_levels' is set to 1,
no adjacent pyramid level exists and the
'novelty_region' is the
novelty region of the first pyramid level.
For debugging, the parameter
'gen_result_handle' can be set to
. This way,
the novelty score images and novelty regions of the individual pyramid
levels are stored in a result handle.
They can then be read with
In general, the novelty score images describe how well each pixel
fits to the texture inspection model that was created within the training
process. The novelty regions of the single pyramid levels are then
calculated by applying the
'novelty_threshold' to the
corresponding novelty score images.
Inspecting the novelty score images and the novelty regions provides you with
hints which kinds of errors are detectable on which pyramid levels. Thus,
you can decide whether parameters for the training and the classification
should be adjusted, e.g., whether it is reasonable to explicitly select only
a subset of pyramid levels using the parameter
The functionality of the texture inspection can be altered by various
parameters. The parameters can be queried with
and altered with calls to
It is possible to write a texture inspection model to file with the
. Thereby, only the texture
inspection model is stored. The images which were possibly added can be
queried with the help of
and stored separately by calling
To help reduce the memory required by the texture inspection model, it is
possible to delete previously added images from the model with
Furthermore, it is possible to serialize and deserialize the texture
inspection model with
In the following, the most important terms that are used in the context of texture inspection are described:
Images that are used for training.
Images which are compared to the trained texture inspection model.
A collection of neighboring pixels.
The pixel values within a patch.
A texture feature that is used for training.
In the test process the texture features of the test
images are compared to the texture inspection model and their
'novelty score' is calculated. The larger this value, the more
probable it is that the individual texture feature does not fit to the
texture inspection model.
The novelty threshold is determined during the training process. Texture features with a novelty score below the novelty threshold fit to the texture inspection model. Texture features with a higher novelty score do not.