METHOD FOR EVENT-BASED SEMANTIC CLASSIFICATION
A method of automatically classifying images in a consumer digital image collection, includes generating an event representation of the image collection; computing global time-based features for each event within the hierarchical event representation; computing content-based features for each image in an event within the hierarchical event representation; combining content-based features for each image in an event to generate event-level content-based features; and using time-based features and content-based features for each event to classify an event into one of a pre-determined set of semantic categories.
Latest Intellectual Ventures Fund 83 LLC Patents:
This application is a Divisional of U.S. application Ser. No. 12/273600, filed Nov. 19, 2008, incorporated herein by reference in its entirety.FIELD
The invention relates generally to the field of digital image processing, and in particular to a method for classifying digital images into semantic categories.BACKGROUND
The proliferation of digital cameras and scanners has led to an explosion of digital images, creating large personal image databases. The organization and retrieval of images and videos is already a problem for the typical consumer. Currently, the length of time spanned by a typical consumer's digital image collection is only a few years. The organization and retrieval problem will continue to grow as the length of time spanned by the average digital image and video collection increases, and automated tools for efficient image indexing and retrieval will be required.
Many methods of image classification based on low-level features such as color and texture have been proposed for use in content-based image retrieval. A survey of low-level content-based techniques (“Content-based Image Retrieval at the End of the Early Years,” A. W. M. Smeulders et al, IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(12), December 2000) provides a comprehensive listing of relevant methods that can be used for content-based image retrieval. The low-level features commonly described include color, local shape characteristics derived from directional color derivatives and scale space representations, image texture, image transform coefficients such as the cosine transform used in JPEG-coding and properties derived from image segmentation such as shape, contour and geometric invariants. Though these features can be efficiently computed and matched reliably, they usually have poor correlation with semantic image content.
There have also been attempts to compute semantic-level features from images. In
WO 01/37131 A2, visual properties of salient image regions are used to classify images. In addition to numerical measurements of visual properties, neural networks are used to classify some of the regions using semantic terms such as “sky” and “skin.” The region-based characteristics of the images in the collection are indexed to make it easy to find other images matching the characteristics of a given query image. U.S. Pat. No. 6,240,424 B1, discloses a method for classifying and querying images using primary objects in the image as a clustering center. Images matching a given unclassified image are found by formulating an appropriate query based on the primary objects in the given image. U.S. Patent Application Publication No. 2003/0195883 A1 computes an image's category from a pre-defined set of possible categories, such as “cityscapes.”
These semantic-level features are also not the way users recall and search for images in their collection. Users' recollection of photographs is often based on the event that was captured. For example, photographs may be identified as “Grand Canyon vacation,” “Mom's birthday party,” “Joe's baseball league” and so on. There are mechanisms available in current software to manually enter such tags or captions to identify photographs. However, a need exists to automate this labor-intensive process, so that a user is able to search by common types of events without having to tag the images first. Further, the user can combine event type with other semantic features such as people present in the image, location or activity to narrow the search to relevant images.SUMMARY
It is an object of the present disclosure to classify images or videos in a digital image collection into one of several event categories. This object is achieved by a method of automatically classifying images in a consumer digital image collection, comprising:
(a) generating an event representation of the image collection;
(b) computing global time-based features for each event within the hierarchical event representation;
(c) computing content-based features for each image in an event within the hierarchical event representation;
(d) combining content-based features for each image in an event to generate event-level content-based features; and
(e) using time-based features and content-based features for each event to classify an event into one of a pre-determined set of semantic categories.
The organization and retrieval of images and videos is a problem for the typical consumer. Automated tools are needed that can understand the content of the images and provide the ability to search the collection using semantic concepts such as events, people and places. The embodiments provide automatic classification of images in a collection into semantic event categories. This will permit the consumer to search for and browse images in the collection depicting specific events. The images and videos are automatically labeled with event category labels that can enable the automated generation of event-specific creative media outputs. The embodiments provide the advantage of permitting users to search for images or videos in the collection that are part of specific event categories. Further, the embodiments have the advantage that images are automatically labeled with event category labels that can enable the automated generation of event-specific creative media outputs.
The present disclosure can be implemented in computer systems as will be well known to those skilled in the art. The invention has been described in detail with particular reference to certain preferred embodiments thereof, but it will be understood that variations and modifications can be effected within the spirit and scope of the invention. Videos in a collection are treated as groups of keyframe images and included in the term “images” in the rest of the description.
The events detected continue to be chronologically ordered in a timeline from earliest to latest. Using the method described above, it is not possible to detect single events that span a long period of time (days) and encompass a variety of activities and settings (for example, a long vacation covering multiple destinations) or events that occur in distinct parts separated by some hours from each other (for example, a sporting event with many matches or a wedding). Gaps in photo-taking corresponding to the overnight period also cause breaks in event continuity. Further processing is needed to detect these super-events, defined as a grouping of multiple contiguous events that may span multiple days.
Table 1 has a list of example features that are collected from available algorithms. These features are found to be strongly correlated with the semantic event category. The first column of the table shows the name of the feature. The second column indicates the levels the feature can have. For example, in the simplest case, a feature can have only two levels—present and not present. The feature detection output can also be quantized into multiple levels to indicate either extent or degree of confidence of the feature. For some features, an experimentally determined threshold is used to test presence e.g. water is present if more than 25% of the image pixels are detected to be water (as shown in Table 1). This ensures that the feature is a significant part of the image. These features are detected at the image level. Image-level feature detection results are combined to obtain event-level content-based features. If a sufficient percentage of images in an event are tagged with the (non-zero) feature level, the event is tagged with that feature level. This threshold percentage is indicated in the third column of Table 1. For example, 15% of the images in an event would need to have a “Present” level for the feature “Grass” for the event to have the level “Present” for the feature “Grass.”
Semantic event categories are pre-determined based on studies of typical consumer picture-taking occasions and cover 80-90% of all images captured. Table 3 shows four top-level semantic event categories and their sub-categories. The features computed by 112 and 114 of
The Bayesian Belief Network is set up as shown in
The accuracy of the classifier may be further improved by consideration of auxiliary factual data (information) directly or indirectly associated with the captured event, illustrated as box 116 in
Such factual information can be applied by the system as part of the BBN.
The BBN is trained on labeled input/output data to calculate the conditional probabilities at each arc of the network. The a priori probabilities (or priors) are also learned from the labeled training data. Referring to
Since conditional probabilities for some of the features in the extended feature set can be difficult or infeasible to obtain through training data, the system can also optionally apply heuristics or rules to validate or refine the output of the BBN of
Rules are also used to further refine event categories to subcategories. For example, the BBN may not be able to significantly distinguish between the various subcategories of party for a particular event, but rules leveraging auxiliary data can make the distinction. Suppose the BBN has determined a top-level category of “Social Gathering.” A rule-based approach can apply the sub-category “Social Gathering-birthday” by using the results of the people recognition algorithm to determine the people portrayed in the event and applying the common sense rule that states that a social gathering is a birthday party if a person portrayed in the event has a birthday at or near the time of the event; information about individual birthdays is part of the data stored in an auxiliary factual database 116 of
The BBN of
Vacation: “Daytrip” if event duration is within the same calendar day, “Weekend” if event falls during the weekend days, “Getaway” if the vacation is long (greater than 4 days), “Holiday” if it falls during a holiday and the location is determined to be outside the user's home area. Sports: “Field” if results from detector for field and “grass” are positive, “Water” if result from detector for water is positive, “Indoor” if result of indoor/outdoor detector is indoor, and “Other” if none of the above detectors show positive result. Family Moments: The category is based on the relationship of the user to the people recognized in the images. Social Gathering: “Wedding” and “Birthday” are based on information from the personal calendar of a user, “Holiday” is based on the calendar and holidays observed in the user's geographic location, and “Party” is based on the presence of many people recognized by the people recognition system as being commonly occurring in the user's collection (i.e. not strangers).
In addition to permitting the user to search and browse by event type, the semantic category of event can be used to author creative output for the user. This semantic information together with other contextual information (including, but not limited to image metadata, date/time, GPS, location, or any combination thereof) can be input into an automatic albuming system for generating themed and personalized album based on the type of event detected. For example, a vacation in Europe will suggest the use of a relevant background design and theme that reflects the cultural and regional characteristic of the location where the event took place. A party event will evoke the use of a fin and whimsical theme and mood for the album. Other custom output products can be created using the event information including collages, cups, T-shirts, and multimedia DVD and CD that includes audio and music. In addition, the semantic event information enables more efficient searching and browsing of the user collection. For instance, the user can easily search by text when the events have been automatically annotated by the system. Yet another application will be for targeted advertising based on the event detected. For example, when an outdoor sporting event is detected, relevant advertising of sporting goods can be targeted to the user.
The invention has been described in detail with particular reference to certain preferred embodiments thereof, but it will be understood that variations and modifications can be effected within the spirit and scope of the invention.PARTS LIST
- 105 Digital image collection
- 110 Hierarchical event detector
- 112 Content-based feature detectors
- 114 Time-based feature detectors
- 116 Auxiliary factual data
- 120 Event classifier
- 122 Event label
- 225 Event and sub-event detector
- 230 Inter-event duration calculator
- 235 Density-based classifier
- 240 Event hierarchy generator
- 305 Event labels based on image features node
- 310 Duration node
- 315 People present node
- 320 Outdoors node
- 405 Event labels based on extended feature set node
- 410 Subject distance node
- 415 GNIS feature class node
- 510 Validation rules
- 515 Refinement rules
- 520 Rule-based classifier
- 530 Feature extractors
- 602 Digital image training collection
- 607 Manual event labeler
- 610 Correlated feature detectors
- 620 Event classifier training
- 710 Probabilistic classifier . . .
1. A method comprising:
- generating, using a processor, time-based event boundaries detected in a plurality of images;
- computing inter-event durations;
- grouping events into clusters based on the inter-event durations; and
- validating, using a rule-based system, that each event belongs to an associated cluster based on event level content based features.
2. The method of claim 1, wherein grouping events into clusters includes using density-based clustering.
3. The method of claim 1, wherein the inter-event durations span multiple days.
4. The method of claim 1, wherein the inter-event durations span small duration gaps.
5. The method of claim 4, wherein the small duration gaps are less than 18 hours.
6. The method of claim 1, wherein validating, using a rule-based system, comprises referencing a database of auxiliary factual information associated with subjects identified in the plurality of images.
7. The method of claim 1, further comprising determining a location for each image in an event and grouping events into clusters based upon the locations.
8. The method of claim 1, further comprising determining a subject distance for each image in an event and grouping events into clusters based upon the determined locations.
9. A system comprising:
- one or more processors configured to: generate time-based event boundaries detected in a plurality of images; compute inter-event durations; group events into clusters based on the inter-event durations; and validate, using a rule-based system, that each event belongs to an associated cluster based on event level content based features.
10. The system of claim 9, wherein events are grouped into clusters using density-based clustering.
11. The system of claim 9, wherein the inter-event durations span multiple days.
12. The system of claim 9, wherein the inter-event durations span less than 18 hours.
13. The system of claim 9, wherein the validation includes referencing a database of auxiliary factual information associated with subjects identified in the plurality of images.
14. The system of claim 9, wherein the one or more processors are further configured to determine a location for each image in an event and group events into clusters based upon the locations.
15. A non-transitory computer-readable medium having instructions stored thereon, the instructions comprising:
- instructions to generate time-based event boundaries detected in a plurality of images;
- instructions to compute inter-event durations;
- instructions to group events into clusters based on the inter-event durations; and
- instructions to validate, using a rule-based system, that each event belongs to an associated super-event cluster based on event level content based features.
16. The non-transitory computer-readable medium of claim 15, wherein events are grouped into clusters using density-based clustering.
17. The non-transitory computer-readable medium of claim 15, wherein the inter-event durations span multiple days.
18. The non-transitory computer-readable medium of claim 15, wherein the inter-event durations span less than 18 hours.
19. The non-transitory computer-readable medium of claim 15, wherein the instructions to validate includes instructions to reference a database of auxiliary factual information associated with subjects identified in the plurality of images.
20. The non-transitory computer-readable medium of claim 15, further comprising instructions to determine a location for each image in an event and group events into clusters based upon the locations.
International Classification: G06F 17/30 (20060101);