Best Practices for the image recognition scenario

A first step to have the best experience is to compose a consistent dataset.

To create the database, let's remind you how our Image recognition works:
The visual search solutions allow to find from a query image identical or similar visuals in an image database. The search is based on object recognition, shape or colour and depends on the content of an image rather than on textual information.

In this part, you will learn how to choose your reference images.

To be sure your images are searchable, it is important that they are validated by LTU with some recognition tests.

How to choose your images?

To be recognized, images must be of good quality and ensure a minimum level of graphism. It means:

  • close to reality and referencing the final product
  • not be over-exposed or under-exposed
  • be well centered
  • with no reflection or artifact
  • not blurred
  • composed by various shapes, with lines and crossing points
  • not only textual
  • have marked contrasts
  • not symmetrical

For a single image it is not necessary to multiply references in different conditions. This does not improve the recognition.

Examples of good quality images:

Danone usecase example

Examples of bad quality images:

What kind of images?

Formats: Any image sent to LTU engine must be in the following formats: PNG, GIF, JPEG, JPG, BMP. You will need to convert your images to one of these formats before sending them to LTU Engine.

Image dimension: We recommend using images of 512 x 512 pixels when adding images. Images data/file size should not be smaller than 50kb

Once the dataset is created and your application is set, it is also important to send queries with the highest quality possible for the recognition.
The quality recommendations you have seen in the last paragraph are also valid for the query. But in the majority of the usecases, you don't have the control of the query because queries are sent by the end users. To understand what is a good query in order to guide your users, let's explain strengths and limitations of our solution.

What are we able to find?

LTU Engine Image Matching technology is able to match despite several types of image transformations, detecting not only the exact same image, but also modified versions of the original image and object matches (photographs of same object).

Image transformations can be broadly divided into several groups:

  • Geometric transformations: Scale changes, rotations, translations, flips and projective transformations.
  • Photometric transformations: Color space conversions, gray level transformations, changes in hue, brightness and contrast.
  • Filtering effects: Noise and smoothing.
  • Structural transformations: Composite images, i.e. images that have been overlaid on top of each other, overlay of text, adding of borders and cropping.
  • Recompression: Due to different compression algorithms (such as jpg) and different image encodings, information can get lost and artifacts may appear.
  • Images derived from mobile

Images are often modified with a combination of the above transformations. The LTU image matching technology is robust to these. The matching technology easily matches the above combination which includes Gray scale, blur, re-encoding, projective transformation and overlay composite transformations.

For more information, consult the whitepaper.

What are the search limitations?

To avoid frustrations when the image is not recognized it is important to know the limitation of our image recognition solution. It can help to improve the quality of the query when it is possible.

  • Structure modification too important: If the query image is too distant from the reference, there may not be enough details to recognize the original image.

  • Repetitive patterns: Because repetitive patterns are really similar, pictures parts could be confused or badly identified.

  • False positives: Sometimes image matching signature will detect the same object or scene, but not the same image.

Since image matching is very prone to detect small common parts in images (such as logos), it can sometimes result in false positives because parts of the images did actually match.

  • Indexing Limitations: Sometimes an image may not be indexed. This is due to either an unknown image format or missing information:
    • i.e. uniform coloured images are rejected.
    • i.e. having no distinct image features may be rejected too as in the image below.

For more information, consult the whitepaper.

What kind of images?

Formats: Any image sent to LTU engine must be in of the following formats: PNG, GIF, JPEG, JPG, BMP You will need to convert your images to one of these formats before sending them to LTU Engine.

Image dimension: We recommend using images of 512 x 512 pixels when searching images Images data/file size should not be smaller than 50kb.

Reducing image size for mobile matching applications: Images can be converted to grayscale.

In some usecases, you cannot get access to query images anonymously sent by end users. So, you cannot control the quality of the images, but it will help you guide the user by indicating what and how to flash.

Communicate on the experience

You must be sure that the user is aware that an experience is available and that he knows what to do.
To have a good communication, here are some recommendations:

  • Clearly indicate on the support that it is connected
Example of Anaik's greeting card
Example of a Belin scholar book Example of a Hachette Routard guide

* Clearly indicate to the user what to flash. Use pictures, text and an easily identifiable icon: phone, camera, a symbol like '+' …

In the print use cases, use always the same icons and text to indicate what to flash. Don't hesitate to add your logo or the logo of your application and to add legend under each image to flash.  Example of a VLAN newspaper

  • In print usecases, remind the information on each connected page.

Put always the instructions by the same way to facilitate the identification of the connected images.

  • Communicate about the experience

- on the social media
- on your website
- by distributing flyers
- …

For a long-term experience, do not hesitate to repeat the communication on several time

  • Choose a revelant message

It is necessary to put forward a beautiful promise of experience in order to push the user to the conversion:
- Surprise
- Gaming
- Exclusive content
- …

Educate the user to the scan gestures

For a better and recurring experience, you will need to explain to your users how to scan and describe them the steps to do it properly. For this:

  • In your custom applications, add some tutorial pages to be loaded at the first launch.
  • Give instructions to let users discover the experience and to know how to flash.
 Example of a Hachette Routard guide
  • Use explicit words as “scan”, “flash” or “take picture”.

Always use the same icons and words to get your users used to the gesture and make them identify that there is something to scan.

  • Add a target zone in your application to help the user to take a picture that is good enough for the recognition.
mire example of MyPack application

For the success of your experience, choose pertinent contents to attract and retain customers. Some advices:

  • Create graphic contents by using colours and background images
  • Use GIF to create dynamic contents
  • Vary media: GIF, videos, web links, audio playlists…
  • To be clear, adapt the visual to the usage
  • Do not add too heavy media. A too long loading time would discourage the users
  • Limit the number of Call to Action in the page to minimise the scroll

Example of graphic content for the 20 minutes newspaper:

It is possible to configure different contents depending on the language of the phone directly from the enrichment platform.