create_class_lut_mlp — Create a look-up table using a multi-layer perceptron to classify byte images.
create_class_lut_mlp generates a look-up table (LUT) ClassLUTHandle using the data of a trained multi-layer perceptron (MLP) MLPHandle to classify multi-channel byte images. By using this MLP-based LUT classifier the operator classify_image_class_mlp of the subsequent classification can be replaced by the operator classify_image_class_lut. The classification gets a major speed-up, because the estimation of the class in every image point is no longer necessary since every possible response of the MLP is stored in the LUT. For the generation of the LUT, the parameters NumInput, Preprocessing, and NumComponents defined in the earlier called operator create_class_mlp are important. In NumInput, the number of image channels the images must have to be classified is defined. By using the Preprocessing (see create_class_mlp) the number of image channels can be transformed to NumComponents. NumComponents defines the length of the feature vector, which the classifier classify_class_mlp handles internally. Because of performance and disk space, the LUT is restricted to be maximal 3-dimensional. Since it replaces the operator classify_class_mlp, NumComponents <= 3 must hold. If there is no preprocessing that reduces the number of image channels (NumInput = NumComponents), all possible pixel values, which can occur in a byte image, are classified with classify_class_mlp. The returned classes are stored in the LUT. If there is a preprocessing that reduces the number of image channels (NumInput > NumComponents), the preprocessing parameters of the MLP are stored in a separate structure of the LUT. To create the LUT, all transformed pixel values are classified with classify_class_mlp. The returned classes are stored in the LUT. Because of the discretization of the LUT, the accuracy of the LUT classifier could become lower than the accuracy of classify_image_class_mlp. With 'bit_depth' and 'class_selection' the accuracy of the classification, the required storage, and the runtime needed to create the LUT can be controlled.
The following parameters of the MLP-based LUT classifier can be set with GenParamNames and GenParamValues:
Number of bits used from the pixels. It controls the storage requirement of the LUT classifier and is bounded by the bit depth of the image ('bit_depth' <= 8). If the bit depth of the LUT is smaller ('bit_depth' < 8), the classes of multiple pixel combinations will be mapped to the same LUT entry, which can result in a lower accuracy for the classification. One of these clusters contains 2^(NumComponents*(8-bit_depth)) pixel combinations, where NumComponents denotes the dimension of the LUT, which is specified in create_class_mlp. For example, for 'bit_depth' = 7, NumComponents = 3, the classes of 8 pixel combinations are mapped in the same LUT entry. The LUT requires at most 2^(NumComponents*bit_depth+2) bytes of storage. For example, for NumComponents = 3, 'bit_depth' = 8 and NumOutput < 16 (specified in create_class_mlp), the LUT requires 8 MB of storage with internal storage optimization. If NumOutput = 1, the LUT requires only 2 MB of storage by using the full bit depth of the LUT. The runtime for the classification in classify_image_class_lut becomes minimal if the LUT fits into the cache. The default value is 8, typical values are [6,7,8]. Restrictions: 'bit_depth' >= 1, 'bit_depth' <= 8.
Method for the class selection for the LUT. Can be modified to control the accuracy and the runtime needed to create the LUT classifier. The value in 'class_selection' is ignored if the bit depth of the LUT is maximal, thus 'bit_depth' = 8 holds. If the bit depth of the LUT is smaller ('bit_depth' < 8), the classes of multiple pixel combinations will be mapped to the same LUT entry. One of these clusters contains 2^(NumComponents*(8-bit_depth)) pixel combinations, where NumComponents denotes the dimension of the LUT, which is specified in create_class_mlp. By choosing 'class_selection' = 'best', the class that appears most often in the cluster is stored in the LUT. For 'class_selection' = 'fast', only one pixel of the cluster, i.e., the pixel with the smallest value (component-wise), is classified. The returned class is stored in the LUT. In this case, the accuracy of the subsequent classification could become lower. On the other hand, the runtime needed to create the LUT can be reduced, which is proportional to the maximal needed storage of the LUT, which is defined with 2^(NumComponents*bit_depth+2). The default value is 'fast', possible values are ['fast', 'best'].
Threshold for the rejection of uncertain classified points of the MLP. The parameter represents a threshold on the probability measure returned by the classification (see classify_class_mlp and evaluate_class_mlp). All pixels having a probability below 'rejection_threshold' are not assigned to any class. The default value is 0.5. Restriction: 'rejection_threshold' >= 0, 'rejection_threshold' <= 1.
Names of the generic parameters that can be adjusted for the LUT classifier creation.
Default value: 
Suggested values: 'bit_depth', 'class_selection', 'rejection_threshold'
Values of the generic parameters that can be adjusted for the LUT classifier creation.
Default value: 
Suggested values: 8, 7, 6, 'fast', 'best'
Handle of the LUT classifier.
If the parameters are valid, the operator create_class_lut_mlp returns the value 2 (H_MSG_TRUE). If necessary an exception is raised.