Download presentation
Presentation is loading. Please wait.
1
Support vector machine concept-dependent active learning for image retrieval Reporter: Francis 2005-7-5
2
2 1. Introduction RF: A query refinement scheme to inform a database of his query concept. Such a query refinement scheme (query- concept learner) is a case of pool-based active learning. In the beginning, the unlabeled pool would be the entire image database.
3
3 1-1 Active learning Traditional: passive learning randomly select k images to training set. Active learning: choosing informative images within the pool to users. Such request is called a pool-query It should choose its next pool-query based upon the past answers to previous pool-queries. Our approach is called SVM active learner.
4
4 1-2 Querying example 1
5
5 2. Support vector machines and version space
6
6 3. Active learning and batch sampling strategies Two steps: Sampling: request user feedbacks to query concept key step of SVM active learner. Learning: to be a better classifier Then return k images farthest from the boundary on the relevant side.
7
7 3-1 Speculative sampling It’s computationally intensive. We use it as a yardstick to measure other active-learning strategies.
8
8 3-2 Batch-simple sampling Choosing h unlabeled instances closest to the hyperplane (between the relevant and the irrelevant instances in the feature space).
9
9 3-3 Angle-diversity sampling For maintaining the diversity. Diversity of samples is measured by angles between the samples: Score: Trade-off parameter is set at 0.5 Unlabeled instance
10
10 3-4 Error-reduction sampling 11 1
11
11 4-1 Concept complexity 1. Scarcity: Using hit-rate to indicate it. Ex: keyword “sun” v.s “sunrise” 2. Diversity: Ex: the “flowers” concept is more diverse than the “red roses” concept.
12
12 4-1 Concept complexity (con.) 3. Isolation: Input space isolation : Keyword isolation Using association-rules mining 1 Ex: fruit apple(0.5) v.s apple fruit(0.7) 1 、 0.25 “Fruit” is poorly isolated from “apple” 2 、 0.21 “Apple” is well isolated from “fruit”
13
13 4-2 Limitations of active learning When the target concept instances are scarce and not well isolated, active learning will be ineffective. 1. Scarce: common situation is that target concept matching images is less than 1% It needs many feedback iterations to obtain positive feedback.
14
14 4-2 Limitations of active learning (con.) 2. Not well isolated:
15
15 4-3 concept-dependent active learning algorithms State C – keyword disambiguation State B – input-space disambiguation State D – key word & space disambiguation
16
16 4-3-1 keyword disambiguation 消去跟負回饋有相同 關鍵字的 unlabeled set 元素 隨機找出 n 個具含糊 關鍵字的元素
17
17 4-3-2 input-space disambiguation
18
18 4-3-3 State D & A State D: using DK and DS algorithms State A: adapt to Diversity Ex: “flowers” concept: learner may need to be more explorative and search for flowers of all colors. Classification score function: In state A, λis reduced to result in more weight in angle diversity during sample selections
19
19 5. Experiments Using five image datasets from Corel image database. Four-category set: 602 images Ten-category set: 1277 images Fifteen-category set: 1920 images 107-category set: 50000 images Large set: 300K image from a stock-photo company.
20
20 5-1 active learning v.s passive 第一輪 20 張 random sampling ,之後 active learning 選 10 張或 20 張
21
21 5-2 against traditional relevance feedback schemes
22
22 5-3 Sampling method evaluation Using 107 category dataset Error reduction sampling
23
23 5-3 Sampling method evaluation 1
24
24 5-4 concept-dependent learning
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.