基于内容的图像检索

1 .Assignment 3 Content-Based Image Retrieval 1

2 .Steps Represent each image in the database we give you by a feature vector. (Preprocess) Design an image distance measure that can compare pairs of image. Retrieve , for each test image, the database images in ascending order of distance to the query. The query itself should have distance 0 and be first. Use the interface provided for retrieval. 2

3 .Initial Processing First apply color clustering to the image to get a labeled image of multiple different cluster labels: 1, 2, 3, ...K. Then apply connected components (provided) to the labeled image to produce a second labeled image that labels each connected component of cluster labels: 1, 2, ... N. A single color cluster may break into more than one component. Possibly perform some noise cleaning to remove small regions. Don’t vary parameters between images. You can get noise cleaning ideas or code from anywhere. 3

4 .Features For each major region (use a size threshold), compute at least the following features: size (number of pixels) given mean color , in RGB, or whatever space you like given a t least the following co-occurrence texture features using spatial relationship d=(1,1): energy, entropy, contrast. centroid (row, column) bounding box (or if you prefer, could be an ellipse) Store the features in the feature vector defined in the code. 4

5 .Extra Credit Features Other region features you want to add RAG (region adjacency graph) including for each pair of adjacent regions: above adjacency below adjacency left adjacency right adjacency other A fancier distance function to handle RAGs. 5

6 .Distance Measure Dist (I 1 ,I 2 ) determines the distance from image I 1 to image I 2 . Compute Dist from a correspondence you find from the regions of I 1 to those of I 2 . Start with a greedy method: for each region of I 1 , find the most similar region of I 2 Do not ask me HOW to do this. That’s for you. 6

7 .More on Distance Measure You should try at least two difference distance measures. They can differ in: attributes used, weights on attributes the actual distance, ie . Euclidean vs. others If you do the graph structure, you need some kind of graph distance. See S&S Section 11.6 or make up your own. 7

8 .Report Turn in a brief report in Word or PDF that describes: the attributes you implemented the distance measures you tried the results of your tests including both the 16 screenshots (1 for each of the 2 distance measures for each of the 8 query images) and your comments. 8

9 .9