A novel log-based relevance feedback technique in content- based image retrieval Reporter: Francis 2005/6/2
2 Outline 1. Introduction 2. Log-based relevance feedback 3. Support vector machines 4. Log-based relevance feedback using SLSVM 5. Experiment results
3 1. Introduction CBIR’s research: 1. Feature analysis and similarity measure: Semantic gap between features and human perceptions. 2. Building the image Indexing with textual descriptions. 3. Relevance feedback
4 2. Log-based relevance feedback Traditional approach: Query expansion (QEX)[6]: It’s good for document retrieval but poor in image retrieval. Log-based relevance feedback: Relevance matrix (RM) Defining correlations between images
5 2. Log-based relevance feedback One given example may be with different relationship value confidence degrees
6 3. Support vector machines Optimization problem formula:
7 4.1 Soft label support vector machine 22
8 4.1 Soft label support vector machine Decision function:
9 4.2 LRF algorithm by SLSVM Training example selection: Using R(i,j) Adding N’ positive and negative training example ranking by S+ and S-.
10 5. Experiment results 1
11 5. Experiment results 1
12 5. Experiment results