Multimedia, Chapter 18: Content-Based Retrieval PDF
Document Details
2003
Li & Drew
Tags
Summary
This chapter from a textbook on fundamentals of multimedia discusses content-based retrieval in digital libraries, detailing several techniques for retrieving images based on features such as color histogram, color layout, and texture. It also explores the concept of video locales and querying on other formats.
Full Transcript
Fundamentals of Multimedia, Chapter 18 Chapter 18 Content-Based Retrieval in Digital Libraries 18.1 How Should We Retrieve Images? 18.2 C-BIRD — A Case Study 18.3 Synopsis of Current Image Search Systems 18.4 Relevance Feedback 18.5 Quantifying Results 18.6 Queryi...
Fundamentals of Multimedia, Chapter 18 Chapter 18 Content-Based Retrieval in Digital Libraries 18.1 How Should We Retrieve Images? 18.2 C-BIRD — A Case Study 18.3 Synopsis of Current Image Search Systems 18.4 Relevance Feedback 18.5 Quantifying Results 18.6 Querying on Videos 18.7 Querying on Other Formats 18.8 Outlook for Content-Based Retrieval 18.9 Further Exploration 1 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.1 How Should We Retrieve Images? Text-based search will do the best job, provided the multi- media database is fully indexed with proper keywords. Most multimedia retrieval schemes, however, have moved to- ward an approach favoring multimedia content itself (“content- based”). Many existing systems retrieve images with the following im- age features and/or their variants: – Color histogram: 3-dimensional array that counts pixels with specific Red, Green, and Blue values in an image. – Color layout: a simple sketch of where in a checkerboard grid cov- ering the image to look for blue skies or orange sunsets, say. – Texture: various texture descriptors, typically based on edges in the image. 2 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.1: How can we best characterize the information con- tent of an image? Courtesy of Museo del Prado. 3 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.2 C-BIRD — A Case Study C-BIRD (Content-Base Image Retrieval from Digital libraries): an image database search engine devised by one of the au- thors of this text. −→ Link to Java applet version of C-BIRD search engine.. C-BIRD GUI: the online image database can be browsed, or searched using a selection of tools: (Fig. 18.2) – Text annotations – Color histogram – Color layout – Texture layout – Illumination Invariance – Object Model 4 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.2: C-BIRD image search GUI. 5 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Color Histogram A color histogram counts pixels with a given pixel value in Red, Green, and Blue (RGB). An example of histogram that has 2563 bins, for images with 8-bit values in each of R, G, B: int hist; // reset to 0 //image is an appropriate struct with byte fields red, green, blue for i=0..(MAXY -1) for j=0..(MAXX -1) { R = image[i][j].red; G = image[i][j].green; B = image[i][j].blue; hist[R][G][B]++; } 6 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Color Histogram (Cont’d) Image search is done by matching feature-vector (here color histogram) for the sample image with feature-vector for im- ages in the database. In C-BIRD, a color histogram is calculated for each target image as a preprocessing step, and then referenced in the database for each user query image. For example, Fig. 18.3 shows that the user has selected a particular image — one of a red flower on a green foliage background. The result obtained, from a database of some 5,000 images, is a set of 60 matching images. 7 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.3: Search by color histogram results. 8 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Histogram Intersection Histogram intersection: The standard measure of similarity used for color histograms: – A color histogram Hi is generated for each image i in the database – feature vector. – The histogram is normalized so that its sum (now a double) equals unity – effectively removes the size of the image. – The histogram is then stored in the database. – Now suppose we select a model image – the new image to match against all possible targets in the database. – Its histogram Hm is intersected with all database image histograms Hi according to the equation X n intersection = min(Hji , Hjm) (18.1) j=1 j – histogram bin, n – total number of bins for each histogram – The closer the intersection value is to 1, the better the images match. 9 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Color Density The scheme used for showing Color Density is displayed in Fig. 18.4. What percentage of the image having any particular color or set of colors is selected by the user, using a color-picker and sliders. User can choose from either conjunction (ANDing) or dis- junction (ORing) a simple color percentage specification. This is a very coarse search method. 10 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.4: Color density query scheme. 11 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Color Layout The user can set up a scheme of how colors should appear in the image, in terms of coarse blocks of color. The user has a choice of four grid sizes: 1 × 1, 2 × 2, 4 × 4 and 8 × 8. Search is specified on one of the grid sizes, and the grid can be filled with any RGB color value or no color value at all to indicate the cell should not be considered. Every database image is partitioned into windows four times, once for every window size. – A clustered color histogram is used inside each window and the five most frequent colors are stored in the database – Position and size for each query cell correspond to the position and size of a window in the image Fig. 18.5 shows how this layout scheme is used. 12 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.5: Color layout grid. 13 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Texture Layout This query allows the user to draw the desired texture dis- tribution. Available textures: zero edge density, medium or high den- sity edges in four directions (0◦, 45◦, 90◦, 135◦) and combina- tions of them. Texture matching is done by classifying textures according to directionality and density (or separation), and evaluating their correspondence to the texture distribution selected by the user. Fig. 18.6 shows how this layout scheme is used. 14 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.6: Texture layout grid. 15 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Texture Analysis Details 1. Edge-based texture histogram A 2-dimensional texture histogram is used based on edge directionality φ, and separation ξ (closely related to repetitiveness). To extract an edge-map for the image, the image is first converted to luminance Y via Y = 0.299R + 0.587G + 0.114B. A Sobel edge operator is applied to the Y -image by sliding the fol- lowing 3 × 3 weighting matrices (convolution masks) over the image. -1 0 1 1 2 1 dx : -2 0 2 dy : 0 0 0 (18.2) -1 0 1 -1 -2 -1 The edge magnitude D and the edge gradient φ are given by q dy D = d2x + d2y , φ = arctan (18.3) dx 16 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Texture Analysis Details (Cont’d) 2. Preparation for creation of texture histogram The edges are thinned by suppressing all but maximum values. If a pixel i with edge gradient φi and edge magnitude Di has a neighbor pixel j along the direction of φi with gradient φj ≈ φi and edge magnitude Dj > Di then pixel i is suppressed to 0. To make a binary edge image, set all pixels with D greater than a threshold value to 1 and all others to 0. For edge separation ξ, for each edge pixel i we measure the distance along its gradient φi to the nearest pixel j having φj ≈ φi within 15◦. If such a pixel j doesn’t exist, then the separation is con- sidered infinite. 17 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Texture Analysis Details (Cont’d) 3. Having created edge directionality and edge separation maps, a 2D texture histogram of ξ versus φ is constructed. The initial histogram size is 193 × 180, where separation value ξ = 193 is reserved for a separation of infinity (as well as any ξ > 192). The histogram is “smoothed” by replacing each pixel with a weighted sum of its neighbors, and then reduced to size 7 × 8, separation value 7 reserved for infinity. Finally, the texture histogram is normalized by dividing by the number of pixels in the image segment. It will then be used for matching. 18 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Search by Illumination Invariance To deal with illumination change from the query image to dif- ferent database images, each color channel band of each im- age is first normalized, and then compressed to a 36-vector. A 2-dimensional color histogram is then created by using the chromaticity, which is the set of band ratios {R, G}/(R + G + B) To further reduce the number of vector components, the DCT coefficients for the smaller histogram are calculated and placed in zigzag order, and then all but 36 components dropped. Matching is performed in the compressed domain by taking the Euclidean distance between two DCT-compressed 36- component feature vectors. Fig. 18.7 shows the results of such a search. 19 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.7: Search with illumination invariance. 20 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Search by Object Model This search type proceeds by the user selecting a thumbnail and clicking the Model tab to enter object selection mode. – An image region can be selected by using primitive shapes such as a rectangle or an ellipse, a magic wand tool that is basically a seed-based flooding algorithm, an active con- tour (a “snake”), or a brush tool where the painted region is selected. – An object is then interactively selected as a portion of the image. – Multiple regions can be dragged to the selection pane, but only the active object in the selection pane will be searched on. A sample object selection is shown in Fig. 18.8. 21 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.8: C-BIRD interface showing object selection using an ellipse primitive. 22 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Details of Search by Object Model 1. The user-selected model image is processed and its features localized (i.e., generate color locales [see below]). 2. Color histogram intersection, based on the reduced chro- maticity histogram, is then applied as a first screen. 3. Estimate the pose (scale, translation, rotation) of the object inside a target image from the database. 4. Verification by intersection of texture histograms, and then a final check using an efficient version of a Generalized Hough Transform for shape verification. 23 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Fig. 18.9: Block diagram of object matching steps. 24 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Model Image and Target Images A possible model image and one of the target images in the database might be as in Fig. 18.10. Fig. 18.10: Model and target images. (a): Sample model image. (b): Sample database image containing the model book. 25 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Image Segmentation vs. Feature Localization Image Segmentation: If R is a segmented region, 1. R is usually connected; all pixels in R are connected (8- connected or 4-connected). 2. Ri ∩ Rj = φ, i 6= j; regions are disjoint. 3. ∪ni=1Ri = I, where I is the entire image; the segmentation is complete. Feature Localization: a coarse localization of image fea- tures based on proximity and compactness – more effective than Image Segmentation. 26 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 1. Locales in Feature Localization Definition: Locale Lf is a local enclosure of feature f. A locale Lf uses blocks of pixels called tiles as its positioning units, and has the following descriptors: 1. Envelope Lf — a set of tiles representing locality of Lf. 2. Geometric parameters — mass M (Lf ) = count of the pix- PM (Lf ) els having feature f, centroid C(Lf ) = i=1 Pi/M (Lf ), PM (Lf ) and eccentricity E(Lf ) = i=1 k Pi − C (L f )k 2/M (L ). f 3. Color, texture, and shape parameters of the locale. For example, locale chromaticity, elongation, and locale tex- ture histogram. 27 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Properties of Locales After a feature localization process the following can be true: 1. ∃f : Lf is not connected. 2. ∃f ∃g : Lf ∩ Lg 6= φ, f 6= g; locales are non-disjoint. 3. ∪f Lf 6= I, non-completeness; not all image pixels are rep- resented. Fig. 18.11 shows a sketch of two locales for color red, and one locale for color blue – The links represent an association with an envelope. Lo- cales do not have to be connected, disjoint or complete, yet colors are still localized. 28 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 L1red L2red L1blue Fig. 18.11: Locales for Feature Localization. 29 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 2. Tile Classification Before locales can be generated, tiles are first classified as having certain features, e.g., red tiles, or red and blue tiles. Dominant color enhancement – Prior to the classification of feature tiles, image pixels are classified as having either dominant color or transitional color. – Dominant colors: pixel colors that do not lie on a slope of color change in their pixel neighborhood. Transitional colors do. – Enhancing the uniformity of the dominant colors is accomplished by smoothing the dominant pixels only, using a 5 × 5 averaging filter, with the proviso that only dominant pixels that have similar color are averaged. 30 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Dominant Color Enhancement (Cont’d) Fig.18.12 shows how dominant color enhancement can clarify the target image in Fig. 18.10. Fig. 18.12: Smoothing using dominant colors. (a): Original image not smoothed. (b): Smoothed image with transitional colors shown in light gray. (c): Smoothed image with transitional colors shown in the replacement dom- inant colors (if possible). Lower row shows detail images. 31 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Tile Feature List Tiles have a tile feature list of all the color features associated with a tile and their geometrical statistics. – On the first pass, dominant pixels are added to the tile feature list – On the second pass, all transitional colors are added to the dominant feature list without modifying the color, yet updating the geometrical statistics – When all pixels have been added to the tiles, the dominant and tran- sitional color feature lists are merged. 32 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 3. Locale Generation Locales are generated using a dynamic 4 × 4 overlapped pyramid linking procedure. (a). The initialization proceeds as: PROCEDURE 18.1 LocalesInit // Pseudo-code begin Let c[nx][ny ] be the 2D array of child nodes. Let p[nx/2][ny/2] be the 2D array of parent nodes. For each child node c[i][j] do Let cn = c[i][j] and pn = p[i/2][j/2]. For each node cnp in the feature list of cn do Find node pnq in the feature list of pn that has similar color. If the merged eccentricity of cnp and pnq has E < τ then Merge cnp and pnq. If pnq doesn’t exist or E >= τ then Add cnp to the start of the feature list of pn. end 33 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (b). After the pyramid linkage initialization, parents compete for linkage with each child: PROCEDURE 18.2 EnvelopeGrowing // Pseudo-code begin Let c[nx][ny ] be the 2D array of child nodes. Let p[nx/2][ny/2] be the 2D array of parent nodes. Repeat until parent-child linkage does not change anymore For each child node c[i][j] do j±1 Let cn = c[i][j] and pn ∈ p[ i±1 2 ][ 2 ] For each node cnp in the feature list of cn do Find node pnq in the feature lists of pn that has similar color and minimizes the distance k C(cnp) − C(pnq )k If the merged eccentricity of cnp and pnq has E < τ then Swap the linkage of cnp to its parent to pnq. Update the associated geometrical statistics. In the parent feature list p remove empty nodes. Go up a level in the pyramid and repeat the procedure end 34 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (c). Following the pyramidal linking, locales having small mass are removed and also sorted according to decreasing mass size for efficiency of search The color update equation for parent locale j and child locale i at iteration k + 1 is T T (k+1) T rj(k) , gj(k) , Ij(k) Mj(k) + ri(k) , gi(k) , Ii(k) Mi(k) rj(k+1) , gj(k+1) , Ij = (18.6) Mj(k) + Mi(k) the update equations for the geometrical statistics are Mj(k+1) = Mj(k) + Mi(k) (18.7) C(k) j Mj (k) + C(k) i Mi (k) C(k+1) j = (18.8) Mj(k+1) 2 2 2 2 (Ej(k) + Cx,j (k) (k) + Cy,j )Mj(k) + (Ei(k) + Cx,i (k) (k) + Cy,i )Mi(k) Ej(k+1) = Mj(k+1) (k+1) 2 (k+1) 2 − Cx,j − Cy,j (18.9) 35 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) (b) Fig. 18.13: Color locales: (a) For the model image. (b) For a database image. 36 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 4. Texture Analysis Every locale is associated with a locale-based texture his- togram. – A locale-dependent threshold is better for generating the edge map. – The threshold is obtained by examining the histogram of the locale edge magnitudes. Global versus locale-based texture measures: The locale-based texture is a more effective measure of tex- ture than is a global one, since the locale-dependent thresh- olds can be adjusted adaptively. – Fig. 18.14 compares a locale-based edge-detection to a global thresh- old based edge-detection. 37 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) (b) Fig. 18.14: Global versus locale-based thresholds. (a) The edge- map for the database image using a global threshold. (b) The edge-map for the database image using a locale-based threshold. 38 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 5. Object Modeling and Matching The object image selected by the user is sent to the server for matching against the locales database. Locale Assignment: the one-to-one correspondence be- tween image locales to be found and model locales. – A locale assignment has to pass several screening tests to verify an object match. – Screening tests are applied in order of increasing complexity and de- pendence on previous tests. The sequence of steps during an object matching process is shown in Fig. 18.9. (a) user object model selection and model feature localization (b) color-based screening test (c) pose estimation (d) texture support (e) shape verification 39 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Object Match Measure Q The object match measure Q is formulated as follows: m X Q=n wiQi (18.10) i=1 n – the number of locales in the assignment m – the number of screening tests considered for the measure Qi – the fitness value of the assignment in screening test i wi – weights that correspond to the importance of the fitness value of each screening test Locales with higher mass (more pixels) statistically have smaller percent- age of localization error. Assignments with many model locales are preferable to few model locales, since the cumulative locale mass is larger and the errors average out. One tries to assign as many locales as possible first, then compute the match measure and check the error using a tight threshold. 40 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Matching Steps The screening tests that are applied to locales in order to generate assignments and validate them are: – Color-based screening tests (Step b): ∗ Illumination Color Covariant Screening ∗ Chromaticity Voting ∗ Elastic Correlation – Estimation of Image Object Pose (Step c) – Texture Support (Step d) – Shape Verification (Step e) – Recovery of Lighting Change 41 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Elastic Correlation Elastic Correlation: the operation that computes the prob- ability that there can be a correct assignment, and returns the set of possible assignments. – Can be used to evaluate the feasibility of having an as- signment of image locales to model locales by using chro- maticity shift parameters. – Having a candidate set of chromaticity shift parameters, each candidate is successively utilized for computing the elastic correlation measure. – If the measure is high enough (higher than 80%, say), then the possible assignments returned by the elastic cor- relation process are tested for object matching using pose estimation, texture support and shape verification. 42 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Elastic Correlation (Cont’d) A A’ E D C g’ B’ C’ B F r’ Fig. 18.15: Elastic correlation in Ω{r 0 , g 0 } Fig. 18.15 shows the elastic correlation process applied in the model chromaticity space Ω{r0, g 0}: – The model image has three locale colors located at A’, B’ and C’. – All the image locale colors, A, B, C, D, E and F, are shifted to the model illuminant. – Elastic correlation is employed, in which the nodes A, B, C are allowed to be located in the vicinity of A’, B’, C’, respectively. 43 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Pose Estimation Method & Texture Support The pose estimation method (Step (c)) uses geometrical re- lationships between locales for establishing pose parameters. – Performed on a feasible locale assignment. – Locale spatial relationships are represented by relationships between their centroids. – Results of pose estimation are both the best pose parameters for an assignment and the minimization objective value. – If the error is within a small threshold, then the pose estimate is accepted. The texture support screening test utilizes a variation of histogram intersection technique, intersecting texture his- tograms of locales in the assignment. – If the intersection measure is higher than a threshold then the texture match is accepted. 44 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Shape Verification Shape verification is by the method of Generalized Hough Transform (GHT): – After performing pose estimation, GHT search reduces to a mere confirmation that the number of votes in a small neighborhood around the reference point is indicative of a match. – The reference point used is the model center since it minimizes voting error caused by errors in edge gradient measurements. Once we have shape verification, the image is reported as a match, and its match measure Q returned, if Q is large enough. – Fig.18.16(a) shows the GHT voting result for searching for a model object (pink book, here) from one of the database images as in Fig. 18.16(b). – Fig.18.16(b) shows the reconstructed edge map for the book. 45 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) (b) Fig. 18.16: Using the GHT for Shape Verification. (a): GHT accumulator array image. (b): Reconstruction of the detected object using the estimated pose and GHT template (edge map). 46 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) Fig. 18.17: Search result for the pink book model with illumina- tion change support. (a): Search results using pose estimation only. 47 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (b) (c) Fig. 18.17 (cont’d): (b): Search results using pose estimation and texture support. (c): Search results using GHT shape verification. 48 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 6. Video Locales Video Locales: a sequence of video frame feature locales that share similar features in the spatio-temporal domain of videos. – Like locales in images, video locales have color, texture, and geometric properties. – They capture motion parameters such as the motion tra- jectory and speed, as well as temporal information such as the life-span of the video locale and its temporal rela- tionships with respect to other video locales. Fig.18.18 shows that while speeding up the generation of locales substantially, very little difference occurs in generation of locales from each image (Intra-frame) and from predicting and then refining the locales (Inter-frame). 49 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) (b) (c) Fig. 18.18: Intra-frame and Inter-frame video locales algorithm results. (a) Original images. (b) Intra-frame results. (c) Inter- frame results. 50 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.3 Synopsis of Current Image Search Systems Some well-known current image search engines are listed here. For URLs and resources, please refer to Further Ex- ploration. −→ Link to Further Exploration for Chapter 18.. – QBIC (Query By Image Content) – UC Santa Barbara Search Engines – Berkeley Digital Library Project – Chabot – Blobworld – Columbia University Image Seekers – Informedia – MetaSEEk – Photobook and FourEyes – MARS – Virage – Viper – Visual RetrievalWare 51 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.4 Relevance Feedback Relevance Feedback: involve the user in a loop, whereby images retrieved are used in further rounds of convergence onto correct returns. Relevance Feedback Approaches – The usual situation: the user identifies images as good, bad, or don’t care, and weighting systems are updated according to this user guidance. – Another approach is to move the query towards positively marked content. – An even more interesting idea is to move every data point in a disciplined way, by warping the space of feature points. 52 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Relevance Feedback (cont’d) The basic advantage of putting the user into the loop by using relevance feedback is that the user need not provide a completely accurate initial query. For a specific example of relevance feedback with respect to the image search engine Mars, please refer to the URL in Further Explorations. 53 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.5 Quantifying Results Precision is the percentage of relevant documents retrieved compared to the number of all the documents retrieved. Desired images returned P recision = (18.13) All retrieved images Recall is the percentage of relevant documents retrieved out of all relevant documents. Desired images returned Recall = (18.13) All desired images These measures are affected by the database size and the amount of similar information in the database, and as well they do not consider fuzzy matching or search result ordering. 54 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.6 Querying on Videos Video indexing can make use of motion as the salient fea- ture of temporally changing images for various types of query. Inverse Hollywood: can we recover the video director’s “flowchart”? – Dividing the video into shots, where each shot consists roughly of the video frames between the on and off clicks of the record button. – Detection of shot boundaries is usually not simple as fade-in, fade- out, dissolve, wipe, etc. may often be involved. 55 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 In dealing with digital video, it is desirable to avoid uncom- pressing MPEG files. – A simple approach to this idea is to uncompress only enough to re- cover just the DC term, thus generating a thumbnail that is 64 times as small as the original. – Once DC frames are obtained from the whole video, many different approaches have been used for finding shot boundaries – based on features such as color, texture, and motion vectors. 56 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Video (Temporal) Segmentation Shots are grouped into scenes — a collection of shots that belong together, and are contiguous in time. Even higher-level semantics exists in so-called film grammar. Audio information is very important for video segmentation. – In a typical scene, the audio has no break within a scene, even though many different shots may be taking place over the course of the scene. Text may indeed be a useful means of delineating shots and scenes, making use of closed-captioning information already available. – Relying on text is unreliable since it may not exist, espe- cially for legacy video. 57 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 Schemes for Organizing and Displaying Storyboards The most straightforward method is to display a 2-dimensional array of keyframes. – Usually a clustering method is used to represent a longer period of time that is more or less the same within the temporal period belong- ing to a single keyframe. Some researchers have suggested using a graph-based method. – A sensible representation of two talking heads might be a digraph with directed arcs taking us from one person to the other, and then back again. Other proxies have also been developed for representing shots and scenes. – Annotation by text or voice, of each set of keyframes in a skimmed video, may be required for sensible understanding of the underlying video. 58 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 An Example of Querying on Video Fig. 18.19: (a) shows a selection of frames from a video of beach activity. Here the keyframes in Fig. 18.19 (b) are selected based mainly on color information (but being careful with respect to the changes incurred by changing illumination conditions when videos are shot). A more difficult problem arises when changes between shots are gradual, and when colors are rather similar overall, as in Fig. 18.20(a). The keyframes in Fig. 18.20(b) are sufficient to show the development of the whole video sequence. 59 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) (b) Fig. 18.19: Digital video and associated keyframes, beach video. (a): Frames from a digital video. (b): Keyframes selected. 60 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 (a) (b) Fig. 18.20: Garden video. (a): Frames from a digital video. (b): Keyframes selected. 61 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.7 Querying on Other Formats A good introduction to using both audio and video cues is the article in text reference. An interesting effort to understand and navigate slides from lectures, based on the time spent on each slide and the speaker’s intonation, is in text reference. A good introduction to search-by-audio is in text reference , and another very interesting approach called “Query- by-Humming” is in text reference. Other features considered for indexing include indexing on ac- tions, indexing concepts and feelings, indexing facial expres- sions, and so on. Clearly, this field is a developing and grow- ing one, particularly because of the advent of the MPEG-7 standard (see Chapter 12). 62 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.8 Outlook for Content-Based Retrieval The present and future trends identified: indexing, search, query, and retrieval of multimedia data based on: 1. Video retrieval using video features: image color and ob- ject shape, video segmentation, video keyframes, scene analysis, structure of objects, motion vectors, optical flow (from Computer Vision), multispectral data, and so-called ‘signatures’ that summarize the data. 2. Use of spatio-temporal queries, such as trajectories. 3. Semantic features; syntactic descriptors. 4. Use of relevance feedback, a well-known technique from information retrieval. 63 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 5. Retrieval using sound, especially spoken documents, e.g. using speaker information. 6. Multimedia database techniques, such as using relational databases of images. 7. Fusion of textual, visual, and speech cues. 8. Automatic and instant video manipulation; user-enabled editing of multimedia databases. 9. Multimedia security, hiding, and authentication techniques such as watermarking. 64 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.8 Outlook for Content-Based Retrieval (Cont’d) Some other directions: – Researchers try to create a search profile so as to encom- pass most instances available, say all “animals”. – Query-based learning: intelligent search engines are used to learn a user’s query concepts through active learning for searches using visual features. – Perceptual similarity measure: focuses on comprehend- ing how images are viewed as similar, by people, on the basis of perception. 65 Li & Drew c Prentice Hall 2003 Fundamentals of Multimedia, Chapter 18 18.9 Further Exploration −→ Link to Further Exploration for Chapter 18. Links to CBR sites are collected in Chapter 18 Further Ex- ploration section of the text website: – A Java applet version of the C-BIRD system utilized in Section 18.2. – A demo of QBIC as an artwork server. – A demo version of the Alexandria Digital Library. – A demo of the Berkeley Digital Library Project. – A demo version of VisualSEEk. – The Informedia project. – A demo of the NETRA system. – A demo of Photobook. – A demo of the Visual RetrievalWare system. – A video describing the technology for the Virage system. – A demo site for the VIPER system. – Links to standard sets of digital images and videos, for testing of retrieval and video segmentation programs. – A link to the keyframe production method for Figs. 18.19, 18.20. 66 Li & Drew c Prentice Hall 2003