Emerging Technologies

Image Classification vs Pattern Recognition vs Object Detection vs Object Tracking–A Primer

December 24, 2018 Algorithms, Artificial Intelligence(AI), Computer Vision, Emerging Technologies, Image Classification, Image Recognition, Object Based Image Analysis, Object Classification, Object Detection, Object Recognition, Object Tracking, Pixel Based Image Analysis No comments

It is a common question that has been asked in all Artificial Intelligence Conference or Discussion Forums. Based on my knowledge, I thought of answering some of these questions:

1.) Image Classification (also called Image Recognition): is the process of creating a thematic image where each pixel is assigned a number representing a class / tag (this also includes ‘unclassified’. In a persons image the classes can be “complexion”, “gender”, “ethnicity”, “representational names”, etc.  

image

2.) Pattern Recognition (also called Object Recognition):  Identifying/recognising the things or objects in an image. This answers the query – what objects or things are depicted in this image? For example, if you are searching for farm lands in an areal image, and other objects such as tractors, sheep’s, cattle’s etc.  There are two classifiers you can familiarize for pattern recognition.  Pixel based classification and object based classification.

  • Pixel based classification: Most of the image analysis algorithms you would encounter would be based on pixel based, However, there is a major drawback for pixel based algorithms that they do not take contextual information in to consideration. Context is how these pixel/object is associated to its environment(a relationship).
  • Object based classification: To overcome this limitation of pixel based classification, we could use Object based image analysis(OBIA) to divide images in to meaningful image-objects and evaluating their characteristics in spatial, spectral and temporal aspects.

Here is how both classification will look like one vs the other:

          image

3.) Object Detection: is another confusing terminology, Object Recognition was able to recognize – what type of object it is? Now object detection answers the query – where is this specific object?  Object detection is based on the point of interest of any given image; for instance electronic devices such as laptop/mobile phone in picture and recognition talked about the specific information about electronic devices, like name, type and other characteristic of particular interest point.

image

4.) Object Tracking: Done in motion pictures like in animated gifs or videos, we want to track how an object is moving, where is it going, or its speed. For example: vehicle tracking in a traffic camera system.Or a Realtime object or people tracking like in Object Detection figure.

image

image

Azure Cognitive Services–Experience Image Recognition using Custom Vision (Build an Harrison Ford Classifier)

December 23, 2018 Algorithms, Artificial Intelligence(AI), Azure AI, Cognitive Services, Compuer Vision Service, Computer Vision API, Custom Vision API, Custom Vision Service, Emerging Technologies, Machine Learning(ML) No comments

Custom Vision Service as part of Azure Cognitive Services landscape of pretrained API services, provides you an ability to customize the state-of-the-art Computer Vision models for your specific use case.

Using custom vision service you can upload set of images of your choice and categorize them accordingly using tags/categories and automatically train the image recognition classifiers to learn from these images and come up with image recognition predictions when you supply an input image. Later consume this service as an API in your existing applications.

image

For example:  Here is how an image of Hollywood Actor – Harrison ford being accurately predicted by the custom model through training using a series of pictures of Harrison Ford through different ages and shapes.

I build this sample during Global AI Bootcamp Letterkenny– Hands on Labs, and will take you further through this article. Harrison Ford is my all time favourite actor.

image

Another example, Harrison Ford was one among 3 in a photo. Here is how the results would look like.

image

Here is how Harrison Ford’ sons picture is being predicted as Harrison Ford, due to similar facial characteristics. |f we further train this model, we can improve it’s capabilities to come up with accurate predictions.

image

Now let us see, how it was implemented.

In this article I am going to use a set of Harrison Ford images found on Google Images and then upload them to Custom Vision service like below. For more accuracy, I tried to collect images of Harrison Ford through different stages of his life, so that computer vision model could evolve to predict more accurate results.

image

Getting Started with Custom Vision:

The Azure Custom Vision API is a cognitive service that lets you build, deploy and improve custom image classifiers. An image classifier is an AI service that sorts images into classes (tags) according to certain characteristics. Unlike the Computer Vision service, Custom Vision allows you to create your own classifications. The Custom Vision service uses a machine learning algorithm to classify images.

Classification and object detection

Custom Vision functionality can be divided into two features. Image classification assigns a distribution of classifications to each image. Object detection is similar, but it also returns the coordinates in the image where the applied tags can be found.

To get started with our example,  first you need to have a Microsoft Account and Register/Login to https://www.customvision.ai

There going to be five steps of activities we are going to do:

1. Setup a Custom Vision Project

Create a new Project by selecting ‘New Project’ button

image

Specify the naming as the followed:

  • Name: HarrisonFordClassifier
  • Description: HarrisonFordClassifier
  • Resource Group: Leave it default to ‘Limited Trial’
  • Project Types: Classification
  • Classification Types: Multi Label (this is essential, we are going to add multiple tags per image in this example: for say ‘Actor’, ‘Person’ and ‘Harrison Ford’
  • Domain: General (for now)

image

2. Upload the Images

a.) Prepare Images

I have gathered a set of images you can download it from here, and extract the HF-Demo-Images.zip in to a folder of your choice.

There are two folders in it  first folder(harrisonford) contains all reference images for training the model and second folder(hf-quicktest) contains all the quick test images we are going to use for evaluating the model.

image

b.) Create Tags

Select ‘+’ icon to create a new tag and create the following tags

  • Actor
  • Hollywood
  • Harrison Ford
  • Person
  • Male

image

Enter Tag Name and click on ‘Save

image

c.) Upload Images

Now that we created all the tags, lets upload the images and tag them with respective tags.

Click on ‘Add Images’ button and select the images from “harrisonford” folder to upload.

image

image


d.) Assign Tags

Now specify the associated tags in My Tags section, selecting from the drop down

image

Then click on Upload

image

Have a review of the images uploaded

image

3. Train

Now let us train the model by selecting the green train button on top right hand side of the page

image

This initiates the first automatic training(Iteration 1) based on the tags you assigned and images associated to it.

image

Once that step is completed let us review the output of the training.

It shows a precision and Recall of 100% indicates our image classification model is trained now to provide Precision of 100% and Recall of 100%.

PS: Recall means out of the tags which should be predicted correctly, what percentage did our model correctly find?

image

4. Evaluate the Model

Now that our classifier is trained, let us evaluate the accuracy. For that we are going to use the sample images from “hf-quicktest” folder.

a.) First click on Quick Test button on top – image

b.) Select a local image or select an image URL

image

image

image

Lets try another image

image

image

Next let us try to upload an image of Ben Ford (Harrison Ford’s son)

image

5. Active Learning

Now that we have couple of accurate predictions, Active Learning involves training the model again from the prediction samples we used. This would make the model evolve to provide us more accurate predictions, for example we correcting the model as it identified that Ben Ford also as Harrison Ford based on similar facial features. In real world, he is a different entity other than his father.

Ben Ford is a Chef by profession. So I am going to upload some of his pictures and tag them as Ben Ford. Also couple of images of both father and son together, and then initiate the training again. Hope they would not feel agitated.

image

image

Now if you look at training performance, Precision and Recall values came down a bit, we can realize it is because we have two persons being tagged with some common tags etc.

image

Let us do a Quick Test with the previous image of Ben Ford again. voilà!, we have some accurate prediction.

image

Similarly, we can repurpose some of the previous prediction images from Predictions tab and add them with right Tags. Then retrain the model again to evolve the model.

image

image

image

The End:

Now that you have learned how you can train Custom Vision API with set of images and retrain them again for more accuracy. Once your training is completed and you are happy with the performance, you can integrate the logic in to your existing apps using Custom Vision REST APIs. You can follow the HOL that covers the integration topic here.

Custom Vision Services provides you state-of-the-art Classification and object detection capabilities to customize it for your specific need with quick and easy steps. This help you reduce your time to market and increase ROI (Return of Investment) for your product lines or ideas.

Start learning today using the below reference links.

References:

Disclaimer: All the images referenced in this article are available on the public domain and there is no way any private images are been included in this examples. We respect Harrison Ford and his family privacy, this article is just an attempt to prove the capabilities of Azure Custom Vision Services, no way intended to insult or invade Mr.Harrison Ford’s privacy.I am a big fan of you sir.

Azure Cosmos DB – TTL (Time to Live) – Reference Usecase

October 9, 2018 .NET, .NET Core, .NET Framework, Analytics, Architecture, Azure, Azure, Azure Cosmos DB, Azure Functions, Azure IoT Suite, Cloud Computing, Cold Path Analytics, CosmosDB, Emerging Technologies, Hot Path Analytics, Intelligent Cloud, Intelligent Edge, IoT Edge, IoT Hub, Microsoft, Realtime Analytics, Visual Studio 2017, VisualStudio, VS2017, Windows No comments

TTL capability within Azure Cosmos DB is a live saver, as it would take necessary steps to purge redudent data based on the configurations you may. 

Let us think in terms of an Industrial IoT scenario, devices can produce vast amounts of telemetry information, logs and user session information that is only useful until we operate on them and take action on them, to be specific up to finate period of time. Once that data becomes surplus, we need an application logic that purges these old records.

With the “Time to Live” or TTL, Microsoft Cosmos DB provides an ability to have your documents automatically purged from database storage after a certian period if time(which you configured)

  • This TTL by default can be set on a document collection level and later can be overridden on a per document basis.
  • Once the TTL is set, Cosmos DB service will automatically remove the documents when its lifetime is over.
  • Inorder to track TTL, Cosmos DB uses an offset field to check when it was last modified.  This field is identifiable as “_ts”, which exists in every document you create.  Basically it is a UNIX epoch timestamp. This field is updated everytime when the document is modified. [Ref: Picture1]

image

[Picture1]

Enabling TTL on Cosmos DB Collection:

You can enable TTL on a Cosmos DB collection simply by using Azure Portal –> Cosmos DB collection setting for existing or during creation of  a new collection)

TTL value needs to be set in seconds – if you need 90 days => 60 sec * 60 min * 24 hour * 90 days = 7776000 seconds

image

[Picture2]

Below is a one of the reference architecture in which Cosmos DB – TTL would be essentially useful and viable to any Iot business case:

image

[Picture3]

Hope that was helpful to get some understanding. For more references visit:  Cosmos DB Documentation

Azure Database for MariaDB: Public Preview

October 4, 2018 Azure Database for MariaDB, Managed Services, MariaDB, OpenSource No comments

During Ignite 2018, Microsoft has announced the availability of Maria DB support in Azure Database services. Today it has been opened for Public Preview for all Azure customers.

mariadbhero

What is MariaDB?

MariaDB is a community-developed fork of the MySQL relational database management system intended to remain free under the GNU GPL.Development is led by some of the original developers of MySQL, who forked it due to concerns over its acquisition by Oracle Corporation.Wikipedia

Azure Database for MariaDB: Public Preview Availability

The Azure Database for MariaDB service is now available in preview. It offers an enterprise-ready, fully managed database service that based on the Community Edition of MariaDB.

The service features open-source compatibility, built-in high availability, dynamic scaling, and flexible pricing. Customers can lift and shift to the cloud and use languages and frameworks of their choice, leveraging the power of MariaDB running on Azure.

To learn more about the service, view the service page, pricing, and documentation.

You can create a MariaDB server by using the Azure portal or Azure CLI.

More References:

NDepend–VSTS/Azure DevOps Integration–Part 01

September 30, 2018 .NET, .NET Core, .NET Framework, Azure DevOps, Best Practices, Code Analysis, Code Quality, Dynamic Analysis, Emerging Technologies, Microsoft, Static Analysis, Tools No comments

In my previous article I wrote an introductory about NDepend and how it will be useful for Agile Team to ensure code quality.

In that article we found how we can use NDepend in a developer machine. Now with this article we will familiarize ourselves in using NDepend in your build automation pipeline in your VSTS/Azure DevOps Build Agent.

There are two types of integration possible for NDepend:

  1. Directly using NDepend Package Extension from VSTS Marketplace
  2. Manual Integration using NDepend Command Line Tool. (This would provide you more control over licensing by setting up the license in your own OnPrem VSTS Build Agent.

For the interest of this article I will cover the use of VSTS Package Extension and using NDepend Build Task in VSTS Build Pipeline.

Installation of NDepend Extension for VSTS/Azure DevOps :

1.) Got to Azure DevOps Market Place:  https://marketplace.visualstudio.com/items?itemName=ndepend.ndependextension

image

2.) Click on Get to Install this extension in to your AzureDevOps account and follow the steps. For the demo purpose I am starting with 30 day free trial, otherwise you can go ahead and buy the full license.

image

image

image

3.) Now when you get back to Azure DevOps project, you can see the NDepend side menu enabled, this is where you would see the report summary of your project.

image

Integration NDepend into Azure DevOps Pipeline :

1.) Select “NDepend Task” and add in to Pipeline

image

image

Note:

  • You can choose to stop the build when at least one quality gate fails.
  • You also need to specify the NDepend project file customized for your project, otherwise NDepend will use their default project file configuration.  Having your own NDepend project file will provide you more control over the policies for the scan.

Queue a new Build and wait for Build to complete. Now you can see the BuildArtifacts includes all NDepend report file.

image

Now you go back to NDepend menu from Left side menu item in Summary Tab. This will provide you detailed view of Technical Debt in your project.

image

image

image

image

image

In the next article I will cover the manual integration steps.

New Microsoft Azure Certifications

September 16, 2018 Azure, Azure SDK, Azure Tools, Certification, Emerging Technologies, MCP, Microsoft, Microsoft Learning, Windows Azure Development No comments

Microsoft has recently announced new certification exam tracks for Azure Administrators, Developers and Architects. Here are the line ups that should help you move your career with right certifications. 

The three new Microsoft Azure Certifications are:

  • Microsoft Certified Azure Developer
  • Microsoft Certified Azure Administrator
  • Microsoft Certified Azure Architect

These certifications would essentially split the previous MCSA/MCSE: Cloud Platform and Infrastructure track and introduce new exams for individual certification track. 

So far I only have limited information available about all the exam numbers for each individual track, as recently Microsoft has made BETA exams available for Microsoft Certified Azure Administrator track. 

These exams are still in BETA, would commence general availability in coming months.  Will keep you posted about newer exams for other tracks as we get to know more. 

References: https://www.microsoft.com/en-us/learning/exam-list.aspx