Organizations that deal with digital assets increasingly use taxonomy-controlled vocabulary in asset metadata. Essentially, it includes a list of keywords that employees, partners, and customers commonly use to refer to and search for their digital assets. Tagging assets with taxonomy-controlled vocabulary ensures that the assets can be easily identified and retrieved in searches.
Compared to natural language vocabularies, tagging based on business taxonomy helps align the assets with a company’s business and ensures that the most relevant assets appear in searches. For example, a car manufacturer can tag car images with model names so only relevant images are displayed when searched to design a promotion campaign.
In the background, the functionality uses the artificially intelligent framework of Adobe Sensei to train its image recognition algorithm on your tag structure and business taxonomy. This content intelligence is then used to apply relevant tags on a different set of assets. Experience Manager Assets automatically applies smart tags to uploaded assets, by default.
You can tag the following types of assets:
|Images (MIME types)||Text-based assets (file formats)||Video assets (file formats and codecs)|
|image/png||DOCX||MOV (H264/AVC, Motion JPEG)|
|image/gif||FLV (H264/AVC, vp6f)|
Experience Manager auto-adds the Smart Tags to the text-based assets and to videos by default. To auto-add Smart Tags to images, complete the following tasks.
A tag model is a group of related tags that are associated with various visual aspects of images being tagged. Tags relate with the distinctly different visual aspects of images so that when applied, the tags help in searching for specific types of images. For example, a shoes collection can have different tags but all the tags are related to shoes and can belong to the same tag model. When applied, the tags help find different types of shoes, say for example by design or by usage. To understand the content representation of a training model in Experience Manager, visualize a training model as a top-level entity comprised of a group of manually added tags and example images for each tag. Each tag can be exclusively applied to an image.
Before you create a tag model and train the service, identify a set of unique tags that best describe the objects in the images in the context of your business. Ensure that the assets in your curated set conform to the training guidelines.
Ensure that the images in the training set conform to the following guidelines:
Quantity and size: Minimum 10 images and maximum 50 images per tag.
Coherence: Ensure that the images for a tag are visually similar. It is best to add the tags about the same visual aspects (such as the same type of objects in an image) together into a single tag model. For example, it is not a good idea to tag all of these images as
my-party (for training) because they are not visually similar.
Coverage: There should be sufficient variety in the images in the training. The idea is to supply a few but reasonably diverse examples so that Experience Manager learns to focus on the right things. If you’re applying the same tag on visually dissimilar images, include at least five examples of each kind. For example, for the tag model-down-pose, include more training images similar to the highlighted image below for the service to identify similar images more accurately during tagging.
Distraction/obstruction: The service trains better on images that have less distraction (prominent backgrounds, unrelated accompaniments, such as objects/persons with the main subject). For example, for the tag casual-shoe, the second image is not a good training candidate.
Completeness: If an image qualifies for more than one tag, add all applicable tags before including the image for training. For example, for tags, such as raincoat and model-side-view, add both the tags on the eligible asset before including it for training.
Number of tags: Adobe recommends that you train a model using at least two distinct tags and at least ten different images for each tag. In a single tag model, do not add more than 50 tags.
Number of examples: For each tag, add at least ten examples. However, Adobe recommends about 30 examples. A maximum of 50 examples per tag are supported.
Prevent false positives and conflicts: Adobe recommends creating a single tag model for a single visual aspect. Structure the tag models in a way that avoids overlapping tags between the models. For example, do not use a common tags like
sneakers in two different tag models names
footwear. The training process overwrites one trained tag model with the other for a common keyword.
Examples: Some more examples for guidance are:
Create a tag model that only includes,
Do not create,
Images used to train: You can use the same images to train different tag models. However, do not associate an image with more than one tag in a tag model. It is possible to tag the same image with different tags belonging to different tag models.
You cannot undo the training. The above guidelines should help you choose good images to train.
To create and train a model for your business-specific tags, follow these steps:
Create the necessary tags and the appropriate tag structure. Upload the relevant images in the DAM repository.
In Experience Manager user interface, access Assets > Smart Tag Training.
Click Create. Provide a Title, Description.
Browse and select the tags from the existing tags in
cq:tags that you want to train the model for. Click Next.
In the Select Assets dialog, click Add Assets against each tag. Search in the DAM repository or browse the repository to select at least 10 and at most 50 images. Select assets and not the folder. Once you’ve selected the images, click Select.
To preview the thumbnails of the selected images, click the accordion in front of a tag. You can modify your selection by clicking Add Assets. Once satisfied with the selection, click Submit. The user interface displays a notification at the bottom of the page indicating that the training is initiated.
Check the status of the training in the Status column for each tag model. Possible statuses are Pending, Trained, and Failed.
Figure: Steps of the training workflow to train tagging model.
To check whether the Smart Tags service is trained on your tags in the training set of assets, review the training workflow report from the Reports console.
All types of supported assets are automatically tagged by Experience Manager Assets when uploaded. Tagging is enabled and works, by default. Experience Manager applies the appropriate tags in near-real-time.
For images and videos, the Smart Tags are based on some visual aspect.
For text-based assets, the efficacy of Smart Tags does not depend on the amount of text in the asset but on the relevant keywords or entities present in the text of the asset. For text-based assets, the Smart Tags are the keywords that appear in the text but the ones that best describe the asset. For supported assets, Experience Manager already extracts the text, which is then indexed and is used to search for the assets. However, Smart Tags based on keywords in the text provide a dedicated, structured, and higher priority search facet. The latter helps improve asset discovery as compared to a search index.
You can curate smart tags to remove any inaccurate tags that may have been assigned to your brand assets, so that only the most relevant tags are displayed.
Moderating smart tags also helps refine tag-based searches for assets by ensuring that your assets appear in search results for the most relevant tags. Essentially, it helps eliminate the chances of unrelated assets from showing up in search results.
You can also assign a higher rank to a tag to increase the tag’s relevance for the asset. Promoting a tag for an asset increases the chances of the asset appearing in search results when a search is performed based on the particular tag.
To moderate the smart tags of your digital assets:
In the search field, search for digital assets based on a tag.
To identify the digital assets that you do not find relevant to your search, inspect the search results.
Select an asset, and then select from the toolbar.
From the Manage Tags page, inspect the tags. If you do not want the asset to be searched based on a specific tag, then select the tag and select from the toolbar. Alternatively, select
X symbol next to the label.
To assign a higher rank to a tag, select the tag and select from the toolbar. The tag you promote is moved to the Tags section.
Select Save and then select OK to close the Success dialog.
Navigate to the Properties page for the asset. Observe that the tag you promoted is assigned a high relevance and, therefore, appears higher in the search results.
By default, Experience Manager search combines the search terms with an
AND clause. Using smart tags does not change this default behavior. Using smart tags adds an
OR clause to find any of the search terms in the applied smart tags. For example, consider searching for
woman running. Assets with just
woman or just
running keyword in the metadata do not appear in the search results by default. However, an asset tagged with either
running using smart tags appears in such a search query. So the search results are a combination of,
running keywords in the metadata.
Assets smart tagged with either of the keywords.
The search results that match all search terms in metadata fields are displayed first, followed by the search results that match any of the search terms in the smart tags. In the above example, the approximate order of display of search results is:
woman runningin the various metadata fields.
woman runningin smart tags.
runningin smart tags.
Enhanced smart tagging is based on learning models of images and their tags. These models are not always perfect at identifying tags. The current version of the Smart Tags has the following limitations:
Inability to recognize subtle differences in images. For example, slim-fit versus regular-fit shirts.
Inability to identify tags based on tiny patterns or parts of an image. For example, logos on shirts.
Tagging is supported in the languages that Experience Manager supports. For a list of languages, see Smart Content Service release notes.
The tags that are not handled relate to:
To train the model, use the most appropriate images. The training cannot be reverted or training model cannot be removed. Your tagging accuracy depends on the current training, so do it carefully.
To search for files with smart tags (regular or enhanced), use the Assets search (full-text search). There is no separate search predicate for smart tags.
The ability of the Smart Tags to train on your tags and apply them on other images depends on the quality of images you use for training.
For best results, Adobe recommends that you use visually similar images to train the service for each tag.