The main thrust of this monograph is related to the problem of the expression of queries by users, who are not usually able to provide visual examples or to express their graphic query in a form appropriate for automatic interpretation. On the other hand, users may be able to indicate examples of images that are similar (or dissimilar) to what they are looking for. This form of providing information about the query is called Relevance Feedback (RF). A visual search engine should, therefore, include an algorithm capable of generalizing a user’s preferences based on his assessment of the relevance of sample images. For applications where image retrieval can be based on low-level features of images, such as colour, texture and shape, we have proposed an algorithm with an approximation of user preferences using an RBF neural network. This approach is adequate for searching in a specified class of images that contain single objects distinct from the background. An example application is the interactive atlas of species that we have developed to test our method.
Główny nurt niniejszej monografii dotyczy problemu formułowania zapytania przez użytkowników, którzy zazwyczaj nie są w stanie dostarczyć graficznych przykładów poszukiwanych obrazów ani wyrazić zapytania graficznego w formie zdatnej do automatycznej interpretacji. Tak więc wyszukiwarka graficzna powinna zawierać algorytm uogólniający preferencje użytkownika na podstawie określenia przez niego, w jakim stopniu przykładowe obrazy spełniają jego oczekiwania (ang. relevance feedback). Dla zastosowań w których wyszukiwanie może być oparte na cechach niskiego poziomu, jak kolor, tekstura czy kształt, zaproponowaliśmy algorytm z aproksymacją preferencji użytkownika przy użyciu sieci neuronowej typu RBF. Metoda ta przewidziana jest do wyszukiwania obrazów należących do określonej klasy, zawierających pojedyncze obiekty wyodrębnione z tła. Przykładowym zastosowaniem jest interaktywny atlas gatunków, który powstał w celu przetestowania zaproponowanej metody.
Wydawnictwa nie prowadzą sprzedaży książek z serii "Rozprawy Monografie". Zainteresowanych prosimy o kontakt z ich autorami.
- Contents
-
Summary 9
Streszczenie 11
Acknowledgments 13
Notation 15
Preface 17
1. State of the art in Content-Based Image Retrieval 20
1.1. Surface descriptors 20
1.1.1. Colour descriptors 21
1.1.2. Ensuring colour constancy in colour-based retrieval 22
1.1.3. Texture descriptors 23
1.1.4. Similarity measures for surface descriptors 25
1.2. Keypoint descriptors and bag-of-features approach 25
1.3. Shape-based retrieval 27
1.3.1. Methods for shape-based retrieval – 2D approach 27
1.3.1.1. Basic region-based features 27
1.3.1.2. Curvature scale space 29
1.3.1.3. Retrieval by alignment 29
1.3.2. Methods for shape-based retrieval – 3D approach 30
1.3.2.1. Feature-based methods 31
1.3.2.2. Graph-based methods 32
1.3.2.3. Geometry-based methods 33
1.3.2.4. Creating 3D models from real objects 33
1.4. Relevance feedback: virtual query and distance-based retrieval 35
1.4.1. Interaction with the user and virtual queries 35
1.4.2. Rocchio formula 36
1.4.3. Combining features in the virtual query scheme 37
1.4.4. Shortcomings of CBIR based on virtual queries 38
2. Shape retrieval by alignment 39
2.1. Introduction to the Hausdorff distance 39
2.2. Formulation of recognition and retrieval as a multilevel optimisation problem 41
2.3. Efficient calculation of the Hausdorff distance 42
2.3.1. State of the art in increasing efficiency of the HD calculation 43
2.3.2. Contour approximation at the level L1 45
2.3.3. Contour approximation at the level L2 46
2.3.4. Contour pruning 49
2.3.4.1. Mathematical basis of the proposed contour pruning method 49
2.3.4.2. An application to contour recognition 50
2.3.4.3. Efficiency 51
2.3.4.4. Experimental time comparison 52
2.3.5. Optimisation in the transformation space 54
2.3.6. Database navigation and pruning 55
2.3.6.1. Using the triangle inequality for determining the search order 55
2.3.6.2. Using additional image features to increase the efficiency of the HD evaluation algorithms 57
2.3.6.3. Experimental time comparison 57
2.3.7. Summary of inaccuracies for the presented speeding-up methods 58
2.4. A practical application: 3D object retrieval by shape alignment 59
2.4.1. 3D modelling vs. direct silhouette matching 59
2.4.2. Dissimilarity measures 61
2.4.3. Implementation 63
3. Retrieval of non-homogeneous objects with preference approximation in feature spaces 66
3.1. General outline of matching scheme and selection of descriptors 66
3.2. Adaptation of distance-based method for complex objects 68
3.3. Approximation of user preferences by RBF 70
3.4. Experimental comparison of algorithms for non-homogeneous object retrieval 72
3.5. Conclusions 74
4. Elicitation of relevant features based on relational MCDM 75
4.1. Relevance feedback by pairwise comparisons 79
4.1.1. Pairwise comparisons and the Analytic Hierarchy Process 79
4.1.2. Basic AHP algorithm 80
4.1.3. The proposed algorithm for information retrieval 82
4.1.4. A practical application: an image retrieval system 87
4.1.5. Performance of the method 88
4.1.6. Conclusions 91
4.2. Relevance feedback by graph of relations 92
4.2.1. The concept of user criteria retrieval based on graph of relations 92
4.2.2. Criteria elicitation and information retrieval based on ELECTRE methodology 93
4.2.2.1. The ELECTRE III method 94
4.2.2.2. Proposed method for criteria elicitation 96
4.2.3. Application to Content-Based Image Retrieval 99
4.2.4. Performance and efficiency 103
4.3. Relevance feedback by individual assessment 105
4.3.1. Reference sets 105
4.3.2. The proposed algorithm for the criteria selection 106
4.3.3. An example of real-life application and performance assessment 108
4.4. Comparison of the performance and efficiency 113
5. Application of multicriteria image analysis and relevance feedback for the glass melting process control 116
5.1. Preliminary image processing: segmentation and mapping 117
5.1.1. Batch segmentation 119
5.1.2. Lens sediment segmentation 120
5.1.3. Image mapping 120
5.2. Analysis of temperature symmetry 122
5.2.1. Batch blanket asymmetry indicator 122
5.2.2. Calculation of CCT 124
5.2.3. The influence of reversals on temperature distribution asymmetry 126
5.2.4. Areas of batch symmetry and glass symmetry 127
5.2.5. Cross indicators of temperature asymmetry 128
5.2.6. Inferring about process settings from the asymmetry indicators 130
5.2.7. An example of the analysis of melting symmetry 131
5.3. Elicitation of melting criteria based on pairwise comparisons 135
5.3.1. Preliminary image processing and the calculation of potential criteria 136
5.3.2. The backward AHP and its application in the elicitation of relevant criteria 137
5.3.3. Elicitation of relevant parameters 141
5.4. Summary of the results and conclusions 144
6. Final conclusions 147
References 149