Implio comes with a wide range of Image Vision models carefully crafted by our team of Data Scientists and moderation experts. These models allow you to make sense of user-generated images and handle their moderation specifically, based on what they contain.

This page describes all available models and how to use corresponding predictions.

Note that these models need to be enabled before they can be used. To do so, please contact our support team.

Types of model

There are 2 different types of models: classification and detection:

• Classification models yield predictions on the entire image.  
  Here's how they are represented in the Implio user interface:

• Detection models on the other hand detect individual objects/elements on the image.  
  Here's how they are represented in the Implio user interface:

Detection models come with additional features – see the Detection models features section below for more information on how to use their unique features.

List of available models

The following table lists all currently available models, organized by topic.

Each model may yield one or several tags per image. The last column contains an explanation on how these tags should be interpreted:

Making use of predictions in automation rules

This section describes how to leverage the output of the above-listed models.

Image tags

Each of the above-listed models can output one or several image tags, for each of the images contained by an item item submitted to Implio for moderation.

These tags are exposed via the following automation variables:

• $images[0].tags
• $images[1].tags
• …

These tags can then be queried using the CONTAINS operator. For instance, you can determine whether the first item's image contains nudity or is suggestive using the following expression:

$images[0].tags CONTAINS ("nudity", "suggestive")

In addition, the $images.tags variable contains the concatenation of tags across all images of an item.

For instance, the following expression checks whether any of the item's images contains nudity or is suggestive:

$images.tags CONTAINS ("nudity", "suggestive")

Those tags are calibrated for high precision (precision is typically around 90%). In other words, you should get a relatively low proportion of false positives.

Uncertain image tags

In addition, image models will output specific tags when the confidence of the prediction is low.

Those tags are exposed via $images[0].uncertain_tags, $images.uncertain_tags and $images.uncertain_tags automation variables.

They are used in the exact same way as regular (high precision) image tags.  
For instance, the following expression will match lower-precision nudity or suggestive images:

$images[0].uncertain_tags CONTAINS ("nudity", "suggestive")

You should use uncertain tags if you need maximum recall (i.e. catch as many images containing the notion you are looking for) at the expense of precision. In other words, you will get a higher proportion of false positives using uncertain tags compared to normal (high-precision) tags.

Uncertain tags are typically sent for manual review, so moderators can have a closer look at images and determine how corresponding items should be handled.

You can create separate automation rules for normal and uncertain image tags and set different actions for them. For instance, you could automatically reject items with normal tags, and send those with uncertain tags to a manual moderation queue for closer inspection. 

Detection models features

Unlike classification models which predict whether the entire image contains the desired notion, detection models detect individual objects within images.

Additional pieces of information – object count and area – are available via specific automation variables.

Object count

The number of objects detected on images are exposed via the following variables:

Object area

The following variables can be used to match images based on the proportion represented by <tag> object(s) over the total image area.

The area for a given tag object (e.g. a face) is calculated as the number of pixels represented by the object's bounding box, divided by the total number of pixels that the image contains. The result is a float number comprised between 0 and 1. 

Sample rule expressions

For instance, the following expression will match items whose first image has text objects representing over 50% of the total image area:

$images[0].text.area > 0.5

Similarly, the following expression will match items where the largest text object across all item's images represents over 50% of the image where it is found:

$images.text.maxArea > 0.5