Content-based image Retrieval Using Color Feature
Download technical report on color-based image retrieval.
We use four color-based image retrieval methods to compare their performance. They are:
A sample query image is |
The images in my database relevant to the left image are
|
||||
|
|
|
|
|
|
|
|
|
|
|
|
The top 10 images retrieved by individual method are tabulated as follows. Relevant images are marked by *.
1. Color Histogram Method (CH) |
|
|
|
|
|
|
Rank 1 * |
Rank 2 |
Rank 3 |
Rank 4 |
Rank 5 |
Recall = 0.58 Precision = 0.43 |
|
|
|
|
|
|
Rank 6 |
Rank 7 |
Rank 8 |
Rank 9 |
Rank 10 |
Notice only 1 relevant image is in top 10 list. Those top 9 non-relevant images present high similarity
with query image in terms of percentage distribution of colors.
2. Color Coherence Vector Method (CCV) |
|
|
|
|
|
Rank 1 * |
Rank 2 |
Rank 3 |
Rank 4 |
Rank 5 |
|
recall = 0.74 precision = 0.61 |
|
|
|
|
|
|
Rank 6 |
Rank 7 |
Rank 8 |
Rank 9 * |
Rank 10 |
Two relevant images are in top 10. CCV improves the performance than CH method because
the location of color distribution is taken into consideration.
3. CCV with Perceptually Similar Color (PSC) |
|
|
|
|
|
Rank 1 * |
Rank 2 |
Rank 3 * |
Rank 4 |
Rank 5 |
|
recall = 0.79 precision = 0.68 |
|
|
|
|
|
|
Rank 6 * |
Rank 7 |
Rank 8 |
Rank 9 * |
Rank 10 |
Performance is better than CCV since PSC captures the fact that human being are less sensitive to
small change of colors and finds images with perceptually similar colors.
4. CCV with PSC after relevance feedback |
|
|
|
|
|
|
Rank 1 * |
Rank 2 * |
Rank 3 |
Rank 4 |
Rank 5 |
Recall = 0.82 Precision = 0.72 |
|
|
|
|
|
|
Rank 6 |
Rank 7 * |
Rank 8 * |
Rank 9 |
Rank 10 * |
Relevance feedback brings more relevant images to the front of top 10 list. RF refines query
such that the query moves towards the space of the set of relevant images.