Zia Vision — An Overview

How Zia Image Validation Works — The Concept 

Setting up an Image Validation Rule

1. You can set an image validation rule for both standard and custom modules.
   
2. You can only upload images of one object type at a time. This applies to all three (desired, undesired, pre-trained) image types.
   

1. Go to  Setup  >  Zia  >  Vision  >  Get Started .
   If you have created rules before, click  New Image Validation Rule .
   
2. In  Create Image Validation Rule  page, enter the  Rule name  (e.g.Validate washing machines).
   
   
3. Under  Where to validate,
   
   • Select the  module  from the drop-down list (e.g.: Products).
     
   • Select the layout required. (e.g: Standard). 
     
   • Select the image field that needs to be validated. 
     You can either validate the record image, or a custom "image upload" field in a module. 
     Here, make a choice on which image field to be validated. (e.g Product Image). 
   • Set the  criteria , either  All records  or  Selected records  (e.g.Product Category is "Washing Machine"].
     In this case, the image validation rule would be applicable only for "washing machine records". 
4. Under Type of Validation
   

1. Choose  Match the image,  if you would like to validate the image based on a complete match. 
   
2. Choose  Object Detection,  if you would like to validate the image based on a specific object detected in it.

Complete image match

In Image Validation based on complete image match, Zia will consider an entire image as a whole, for validation. For instance, in the case of Zylker Electronics discussed above, complete image match is ideal when you want to prevent someone from uploading an image of a TV or radio instead of a washing machine. After studying washing machines from all angles, Zia will be able to declare any other images as invalid.

If this is your requirement, under the Type of Validation section,

1. Choose Match the image. 
   
2. In Upload Training Data, make sure to upload several images of a generic washing machine. 
   You will be required to upload a ZIP file containing this set of training data. 

Now whenever someone uploads a TV by mistake for a "Washing Machine" category record, then Zia will analyze whether the image uploaded has a "complete match" with the training data that you have uploaded. Once Zia fails to find a match, then you will see that the validation has failed and the image would not be taken into the record, but instead, sent for manual approval. 

Object detection

Zia Object Detection involves scanning a specific part of the image for validation. Consider the same example of Zylker Electronics. Going one step further, the washing machine department (so-to-speak) wishes to make sure someone does not add the image of a  top-load washing  machine to a  front-load record  by mistake. Such mix-ups could happen when images are uploaded in bulk. In this case, they could choose to validate the image based on "object detection". In any image you upload for "front-load" washing machine, Zia will look for the "front-load" portion that it is trained to detect. If it is not present, Zia could declare the image invalid. So here, the entire image is not considered for validation, but only the presence of absence of a specific "object" in it. 

If this is your requirement, under the  Type of Validation  section,

1. Choose  Detect Objects. 
   
2. In  Upload Training Data,  make sure to upload several images of a front-load washing machine. 
   You will be required to upload a ZIP file containing this set of training data. 

In the example above, this Zia Vision setting is applicable for all Front load category washing machines. When a product image for a front load machine is added, Zia will check whether it contains the specific "objects" that it is trained to detect— that is the actual "front-load" portion. Only if the front-load portion (the object) is detected in an image, will Zia validate it. Otherwise, if you are uploading a top-load machine by mistake for a front load record, it will be declared invalid and will wait for your manual approval.

Note: You need the specific object to be crystal clear in every image you upload as part of the training data. If you just upload front-load machines as a whole, the training data could end up being vague. Front-load and top-load could look the same to Zia, if you upload whole images. You must upload images of the specific object in question.

Also, remember that the image validation can be done based on whether the criteria is met or not met. That is, you can also write a rule saying, "validate the image only if this object is not detected".

For instance there could be images of defective products that could escape your attention when they are uploaded in bulk. Classic examples are visible scratches or dents on a TV screen, for instance. You could upload various pictures of these scratches or dents to train Zia on what these defects look like. Then, you can write a rule that says, "validate images only when these (defective) objects are NOT detected".

This way, object detection can be used to define desired as well as undesired images.

Other scenarios where object detection is useful:

1. A real estate company wants to ensure that images of "luxury villas" show a swimming pool. So you would simply attach images of swimming pools as the training data for "object detection". So, Luxury Villa images will be validated only if these swimming pool objects are detected. If they are not detected, Zia will declare the image as invalid.
   
2. A car dealership company wants to keep images of defective cars out of their records. So, closeup pictures of dents and scratches could be added as the training data for object detection. Images will be validated only if these objects are NOT detected in the whole image. If they are detected, the image will be declared invalid by Zia.
   
3. A window cleaning/repair company wants their cleaners to upload pictures of perfect windows after a repair job on a custom image field called "Final Look". To get this done, they could upload close-up images of broken/dirty windows to train Zia on what is NOT desirable. So only if such objects are not detected, will the image be valid. If these scratches/broken glass objects are detected, the image will not be validated by Zia.
   

Once you are done with selection of validation type and uploaded corresponding training data, the next step is to note the action on validation failure. 

1. Under Action on Failure, you have a single action. 
   Whenever an image fails validation, it will be sent for manual approval.
   
   The associated record will be created, however the image will wait under the "My Jobs" module to be approved, rejected or updated manually by your sales teams.

Testing the model

1. Go to Setup > Zia > Vision .
   
2. Select the rule for which you want to test the model.
   
3. Click Test the model.
   
   
   
   
4. Upload the image that needs to be validated by Zia.
   
   
5. Zia will validate the image and deliver the result.
   
   
   

Record and Image approvals from My Jobs

Delete or deactivate a rule

You can delete or deactivate a rule if required. If you chose to approve the records that enter the rule, then upon deleting the rule you will be notified about the records that are awaiting an approval. Similarly, upon deactivating a rule the existing records that are awaiting an approval will be retained as is, however new records will not be allowed to enter the rule.

Enabling permission to configure image validation rule and review images

CRM administrators who have the permission to manage configuration or review images enabled in their profile will be able to create image validation rule and review images respectively. You can give the manage configuration permission to selected users who will only create the rule and not be responsible for reviewing the images and similarly the reviewers can be restricted from viewing or creating the configuration.

Under  Profiles  >  Zia  >  Image Validation , there are two permissions which allow the user to do the following:

1. Manage configuration:  Users with this permission can only create, edit, view, enable, disable or delete a rule. They cannot review the images.
   
2. Manage action:  Users with this permission can review the images and approve or reject them. These users will not be allowed to view or create a rule. 
   

Remember that when you toggle on Image validation for a profile, manage action will be automatically enabled. However, you have to select manage configuration manually. 

  

To give profile permission

1. Go to  Setup  >  Users and Control  >  Security Control  and choose a profile.
   Only admin profiles or its equivalent clone can be selected. 
2. In  Profile , go to  Setup Permissions  and click  Zia .
   
3. Under  Zia , toggle on  Image Validation . 
   Manage action will be enabled automatically. Click to disable. 
4. Select  Manage Configuration  or/and  Manage Action .  
   

Best practices for uploading images

Take a look at the  cases  below that highlight the possible problems and the ways to troubleshoot them:  

Case 1.  A stationary supplier wants to validate images of pens and pencils received through bulk orders from the vendors. They upload the images of these objects as classifiers. However, when bulk orders are received most images of biros (ballpoint pens) are marked invalid. 

Solution:  The training data consists of images of all kinds of pens such as   fountain pens, gel pens, marker pens, and pencils except ballpoint pens. Therefore, though Zia can identify a variety of pens it cannot identify a ballpoint pen and shows it as an invalid image. Hence, in order to get accurate result it is important to  upload all images of all the objects that you want to be classified.

Case 2.  A mobile phone dealer wants to validate images of smart phones. He uploads pictures of all the brands of smartphones he sells, but still some images of mobile phones are shown as invalid. 

Solution:  The dealer uploads front images of all brands of smartphones. However, some images that are taken from other angles are difficult to identify and draw similarity between objects. Therefore, it is essential to upload images of objects from various angles for ease of identification and accuracy of results. 

Case 3.  A distributor of toys and puzzles is confused because images of geometric or shape puzzles are identified as invalid objects though the same type of images have been used as classifiers. 

Solution:  A geometric shape appears similar from every angle thus uploading image of an object from either angles would always give accurate results. However, there must be enough images to draw similarity between objects. Hence, while uploading images of objects that may appear similar from all angles it is necessary to have at least 5 to 6 images (minimum requirement for Zia's image validation) of the object. We recommend to repeat one image 5 times for the system to draw correlation.

Points to remember

1. If you want to upload images from your desktop, you must follow these guidelines to achieve the best results from Zia:
   
2. The images must be in these format: JPG, JPEG, PNG, GIF, BMP, and TIFF.
   
3. The training data should be nearly similar to the data that needs to be validated. That is, images of villas, motorbikes, cars etc. should be clear and easily identifiable for accurate validation.
   
4. In general, the training data should have images from multiple angles, resolutions, and backgrounds for variety.
   
5. Vision models generally cannot recognize patterns that humans cannot. So if a human cannot recognize a pattern by looking at the image for around half to one second, the model probably cannot be trained to do it either.
   
6. You can upload as many images as you want for a better accuracy, but minimum 5 images of a particular category must be uploaded for Zia to validate images. 
   
7. Do not combine images from different categories. Use images that best depict your category.
   
8. The training accuracy score will be impacted if these guidelines are not followed. 
   
9. We recommend you to select record or image approval as the action until you are sure of the results shown by Zia.
   

Note

1. We calculate the accuracy of the image validation rule by evaluating the points outlined in the guidelines. Only if a rule has a training accuracy above 80% the system will allow you apply it.
   
2. In case multiple rules are configured, a record will enter a rule based on the configuration that matches.