1 Introduction
2 Extended and compact structural tensors for feature extraction
2.1 Structural tensor for low features detection
2.2 Extended structural tensor
2.3 Compact structural tensor
2.4 Anisotropic filtering for tensor computations
3 TBISA classification framework
3.1 Data preprocessing
3.1.1 PCA-based reduction
3.1.2 Feature selection
-
exhaustive search—an algorithm which guarantees selecting optimal subset of features but it is time consuming, and
-
feature selection based on genetic algorithm—very effective heuristic method which allows to significantly reduce processing time.
3.1.3 Feature normalisation
3.2 Classification/segmentation
3.3 Training procedure
3.4 Pixel-based segmentation
4 Experiments
4.1 Experimental framework
4.2 Experiment 1: classification algorithms
-
k-Near Neighbours (k-NN) [1], a minimal distance classifier with 3 neighbours. The number of neighbour was found in preliminary tests on selected images.
-
Multiple layer perceptron (NN) [2], a classical multi-layer perceptron trained with standard back-propagation algorithm with number of neurons in layers calculated as follow. In the input layer it was equal to number of features, in output layer equal to number of classes, and in a hidden layers calculated according to the following formula (#attributes+ number of classes)/2. Learning rate and momentum parameters were set to 0.3 and 0.2, respectively.
-
Naive Bayes classifier (NB) [13], a classical model which assumes feature independence and normal density distribution of feature set constituents.
-
Random Tree (RT), algorithm which creates a tree with K randomly selected attributes at each node. There was no post-pruning and no tree size limit. K was set according to \(\log _2(\)#attributes)+1).
-
Support Vector Machine (SVM) [22], trained with sequential minimal optimisation procedure using polynomial kernel. Multi-class problems were solved using pairwise classification.
Picture | Classifiers | ||||
---|---|---|---|---|---|
k-NN | NN | NB | SVM | RT | |
326038 | 0.8776 |
0.8886
| 0.8801 | 0.8820 | 0.8250 |
35058 | 0.9742 |
0.9802
| 0.9738 | 0.9742 | 0.9724 |
41033 | 0.8532 | 0.8516 | 0.8378 |
0.8546
| 0.8020 |
66053 | 0.9007 |
0.9076
| 0.8801 | 0.8022 | 0.8712 |
161062 |
0.9180
| 0.7769 | 0.7625 | 0.8091 | 0.8886 |
69040 | 0.8648 | 0.8739 | 0.7023 |
0.8741
| 0.8221 |
134052 | 0.8174 |
0.8395
| 0.8378 | 0.8226 | 0.7984 |
Rank | 2.6429 |
1.8571
| 3.7143 | 2.5000 | 4.2857 |
Picture | Classifiers | ||||
---|---|---|---|---|---|
k-NN | NN | NB | SVM | RT | |
326038 | 0.9919 |
0.9919
| 0.9921 | 0.9919 | 0.9919 |
35058 |
0.9994
|
0.9994
|
0.9994
|
0.9994
|
0.9994
|
41033 |
0.9878
| 0.9793 | 0.9721 | 0.9651 | 0.9749 |
66053 | 0.9840 |
0.9912
| 0.9842 | 0.9840 | 0.9718 |
161062 | 0.9882 | 0.9883 |
0.9884
| 0.9882 | 0.9769 |
69040 | 0.9872 | 0.9876 | 0.9597 | 0.9498 |
0.9893
|
134052 | 0.9818 |
0.9896
| 0.9435 | 0.9610 | 0.9714 |
Rank |
2.1429
| 2.5714 | 2.5714 | 3.2857 | 2.8571 |
Picture | Classifiers | ||||
---|---|---|---|---|---|
k-NN | NN | NB | SVM | RT | |
326038 | 0.8776 | 0.9249 | 0.9058 | 0.9153 |
0.9322
|
35058 | 0.9742 |
0.9938
| 0.9742 | 0.9912 | 0.9841 |
41033 | 0.8803 | 0.9163 | 0.9171 | 0.8903 |
0.9419
|
66053 | 0.8712 | 0.9124 | 0.7210 | 0.8098 |
0.9459
|
161062 | 0.9444 |
0.9967
| 0.9335 | 0.9797 | 0.9397 |
69040 | 0.8660 |
0.9929
| 0.6578 | 0.9366 | 0.9123 |
134052 | 0.8498 | 0.9492 | 0.9305 | 0.8978 |
0.9732
|
Rank | 4.2857 |
1.7143
| 4.0000 | 3.0714 | 1.9286 |
4.3 Experiment 2: number of classes
Attribute | Picture | Class count | |
---|---|---|---|
2 | 3 | ||
RGB | 66053 |
0.9076
| 0.6179 |
161062 | 0.7769 |
0.9113
| |
134052 |
0.8395
| 0.7825 | |
PCA-CST | 66053 |
0.9912
| 0.9299 |
161062 |
0.9883
| 0.9619 | |
134052 |
0.9896
| 0.9612 | |
EST | 66053 |
0.9124
| 0.7734 |
161062 |
0.9967
| 0.9342 | |
134052 |
0.9492
| 0.8688 |
4.4 Experiment 3: normalisation methods
-
Layer Normalisation (LN)—scaling values of layer (i.e. RGB channels or tensor constituents) to the range between 0 and 1,
-
Feature Vector Normalisation (VN)—normalisation of the length of feature vector to standard value 1.
Attribute | Picture | Normalisation method | |
---|---|---|---|
LN | VN | ||
RGB | 326038 | 0.8817 |
0.8886
|
35058 |
0.9802
| 0.9802 | |
41033 |
0.8577
| 0.8516 | |
66053 |
0.9124
| 0.9076 | |
161062 |
0.8527
| 0.7769 | |
69040 | 0.8723 |
0.8739
| |
134052 | 0.8382 |
0.8395
| |
PCA-CST | 326038 | 0.8765 |
0.9919
|
35058 | 0.9834 |
0.9994
| |
41033 | 0.7331 |
0.9793
| |
66053 | 0.8534 |
0.9912
| |
161062 | 0.8264 |
0.9883
| |
69040 | 0.8488 |
0.9876
| |
134052 | 0.8460 |
0.9896
| |
EST | 326038 | 0.8785 |
0.9249
|
35058 | 0.9713 |
0.9938
| |
41033 | 0.8508 |
0.9163
| |
66053 | 0.8728 |
0.9124
| |
161062 | 0.8671 |
0.9967
| |
69040 | 0.8423 |
0.9929
| |
134052 | 0.8757 |
0.9492
|
4.5 Experiment 4: learning set size
-
Is there any optimal set size, i.e. such, which allows to get “acceptable” segmentation accuracy and its further increasing brings very small improvement, not acceptable due to labelling costs?
-
Does the set size affect segmentation accuracy in the same way regardless of feature vector content? In other words, do we need the same number of samples for effective training classifier based on EST, PCA-CST, and RGB?
Picture | Attribute | Learningset size | ||
---|---|---|---|---|
10 | 50 | 90 | ||
326038 | RGB | 0.8824 | 0.8886 |
0.8903
|
PCA-CST | 0.9919 | 0.9919 |
0.9936
| |
EST | 0.9231 | 0.9249 |
0.9385
| |
35058 | RGB | 0.9742 | 0.9802 |
0.9900
|
PCA-CST |
0.9994
|
0.9994
|
0.9994
| |
EST | 0.9742 |
0.9938
| 0.9935 | |
41033 | RGB | 0.8373 | 0.8516 |
0.8784
|
PCA-CST | 0.9589 | 0.9793 |
0.9882
| |
EST | 0.8298 | 0.9163 |
0.9325
| |
66053 | RGB | 0.8628 | 0.9076 |
0.9149
|
PCA-CST | 0.9840 |
0.9912
| 0.9882 | |
EST | 0.8420 | 0.9124 |
0.9228
| |
161062 | RGB | 0.7670 | 0.7769 |
0.7858
|
PCA-CST | 0.9882 |
0.9883
| 0.9882 | |
EST | 0.8903 | 0.9967 |
0.9971
| |
69040 | RGB | 0.8519 |
0.8739
| 0.8694 |
PCA-CST | 0.9325 | 0.9876 |
0.9896
| |
EST | 0.8828 | 0.9929 |
0.9972
| |
134052 | RGB | 0.7843 | 0.8395 |
0.8515
|
PCA-CST | 0.9610 |
0.9896
| 0.9884 | |
EST | 0.8725 | 0.9492 |
0.9853
|
4.6 Experiment 5: feature vector representation, and feature reduction methods
-
feature selection from extended structural tensor using genetic algorithms (GA-FS-EST),
-
feature selection from extended structural tensor exhaustive search (ES-FS-EST).
Image | Feature set representation | ||||
---|---|---|---|---|---|
RGB | PCA-CST | EST | GA-FS-EST | ES-FS-EST | |
326038 | 0.8871 |
0.9925
| 0.9288 |
0.9925
|
0.9925
|
35058 | 0.9815 |
0.9994
| 0.9872 | 0.9880 | 0.9880 |
41033 | 0.8558 | 0.9755 | 0.8929 |
0.9832
|
0.9832
|
66053 | 0.8951 |
0.9878
| 0.8924 | 0.9152 | 0.9152 |
161062 | 0.7765 | 0.9882 | 0.9614 |
0.9969
|
0.9969
|
69040 | 0.8650 | 0.9699 | 0.9576 |
0.9916
|
0.9916
|
134052 | 0.8251 |
0.9797
| 0.9357 | 0.9629 | 0.9629 |
Rank | 4.8570 | 1.8571 | 3.8571 | 1.6429 | 1.6429 |