train_texture_inspection_model — Train a texture inspection model.
train_texture_inspection_model( : : TextureInspectionModel : )
train_texture_inspection_model trains a texture inspection model
with all training images which were added by
The complete texture inspection process works with image pyramids. The depth of the pyramid is determined by the parameter 'num_levels' (or 'levels' if the levels have been set explicitly). In the first step of the training, texture features are extracted and added to the set of training samples for each pyramid level. For each pyramid level a Gaussian Mixture Model (GMM) classifier is determined with all training samples of the corresponding pyramid level. In a third step, the training samples are used to determine a novelty threshold for each pyramid level. In the following, the three training steps are explained in detail:
Feature extraction extracts a feature for each pixel within
the training images. The texture features are calculated according to the
settings of the 'patch_*' parameters, which can be manipulated
set_texture_inspection_model_param. Each texture feature
is added to the training data of the GMM.
During the training of the GMMs, the optimal 'gmm_*'
parameters are calculated from the training data. The dimension of the
single GMMs is determined by the size of the feature patches. Please refer to
set_texture_inspection_model_param for details. For large patch
sizes, it can be extremely difficult to obtain a good approximation of the
optimal GMM parameters. Also, the run time and memory usage rises
significantly. Therefore we recommend to alter the default patch size with
More information on GMM classifiers can be found in the description of the
The calculation of the novelty thresholds is required
to be able to distinguish between defective and non-defective
texture. First, the novelty score of each training sample is
determined with the GMMs determined in step 2. Then, based on
the resulting novelty scores, novelty thresholds are determined for each
pyramid level. The automatic determination of the novelty thresholds
can be influenced with the parameter 'sensitivity' that can
be set with
After the training was successful it is possible to classify images
apply_texture_inspection_model. Each pixel is
assigned a novelty score that is compared to the novelty
threshold, which was determined within the third step of the training. For
optimal results it is possible to adapt the novelty threshold of the single
pyramid levels with
In general, the manipulation of all parameters, with the exception of
'gen_result_handle' and 'sensitivity', requires a retraining
of the texture inspection model. Depending on which part of the training is
affected by the corresponding parameter, either the whole training or only
single steps have to be recalculated. This can lead to strong variations in
the time required for retraining a texture inspection model. A precise
description which parameters influence which parts of the algorithm are
For an explanation of the concept of the texture inspection see the introduction of chapter Inspection / Texture Inspection.
This operator modifies the state of the following input parameter:
During execution of this operator, access to the value of this parameter must be synchronized if it is used across multiple threads.
TextureInspectionModel(input_control, state is modified) texture_inspection_model
Handle of the texture inspection model.
* Create texture inspection model create_texture_inspection_model ('basic', TextureInspectionModel) * Make this short example fast: set_texture_inspection_model_param (TextureInspectionModel, \ 'gmm_em_max_iter', 1) * Read and add training images read_image (TrainImage, 'carpet/carpet_01') add_texture_inspection_model_image (TrainImage, TextureInspectionModel, \ Indices) * Train the model train_texture_inspection_model (TextureInspectionModel) * Read and apply a test image read_image (TestImage, 'carpet/carpet_02') apply_texture_inspection_model (TestImage, DefectCandidates, \ TextureInspectionModel, \ TextureInspectionResultID)
train_texture_inspection_model returns the value TRUE
if the given parameters are valid and within acceptable range.
Otherwise, an exception will be raised.
X. Xianghua, M. Mirmehdi: “TEXEMS: Texture Exemplars for Defect Detection
on Random Textured Surfaces”; IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 29, No. 8; August 2007.
T. Boettger, M. Ulrich: “Real-time Texture Detection on Textured Surfaces with Compressed Sensing”; Pattern Recognition and Image Analysis, Vol. 26, No. 1, pp. 88-94; January 2016.