Lecture 6

I am Dr. Rajeev Bishra from IIT, Patna. The topic of today's lecture is ML-based image classifier at IoT Edge. In this lecture, we will cover the following topics. Basics of computer vision in machine learning, different techniques of computer vision like image classification, detection, segmentation, object detection models like R-CNN, faster R-CNN, so YOLO, SDD and so on. Then we will look into the computer vision from Azure that is how that is being supported as software as a service SaaS. So let us give an introduction to the computer vision. So as you know that computer vision is a sub-branch of machine learning and this deals with giving the computer the ability to see and interpret and extract the information from the images and the videos. So the videos can be seen as the collection of images. So here in this particular figure, you can see there are different objects which you can see in an image. So how to interpret whether it is a bottle or it is a laptop or some bowl or it is a bag, chair and so on. So this particular ability is quite natural in human beings but how this capability is to be enabled into the machine using the concept which is called computer vision and that can be enabled with the help of machine learning and for that we are going to discuss some of the preliminaries and the basics and then we will see how these capabilities can be enabled into an IoT edge. So let us understand a bit deeper about computer vision, its working. Now here to train the computer vision in a machine learning model, you essentially need to feed hundreds and thousands of images of interest. So for example, if you want to find out in an image whether it is a cat or it is some other animal. So you have to now get such images as an example of a cat or some other animal. So these particular thousands of images will now be used as the data to train the computer vision model. For that, this computer vision model will apply some complex mathematics and feature extraction to find out in contrast what is the background and what is the object into that image. So based on that the model learns the key understanding or the properties of the objects that defines the cat or some other animal and distinguish that this is the basic understanding of the computer vision. And it mimics the way humans can also understand it with the help of their vision and understanding through the neural networks. The only difference here is that now instead of neural network, it becomes an artificial neural network that is being simulated into the machines. But the capabilities are means inspired from human beings. So take this example, you have the rich dataset of thousands of images of a particular object cat or animal and this particular so many number of images of a cat will be fed to a machine learning model that is running that computer vision that is the algorithm. This particular algorithm or a machine learning model for computer vision will is trained is called a trained once it goes through all these particular images and understands the features of this object that is called a cat. So you can see the output that means when a new image is given to the trained model it will be able to detect with some percentage what is there into that image whether it is a dog, it is cat or it is a duck or a bird. These are different classes, they are called a classes. Now you can see that if you apply a threshold that if it is more than 90 percent probability of a particular class then you can declare that this is an image of a dog and the other image is the image of a cat. So this is a simple way to understand how the computer vision with the help of algorithm and the dataset works. So now let us see the other view that is called the data analytics part which is associated with this computer vision. So with every machine learning model the model is not only the important part the fundamental fact is that it is going to determine how good your model is the data which you feed to it. So therefore this particular aspect which is called the data science is to prepare the data or it is also called the dataset and this particular dataset if it is properly prepared by the domain experts and with the help of a data scientists then your model will be performing a very good. So therefore there is another point that we have to focus on the fact that your model is only as good as your particular data it is not that machine learning model is doing good or bad it is the dataset which becomes very important factor to train the machine learning model and this aspect is called the data science. So to perform this computer vision you have to have a good understanding of a data science and those experts are called data scientists. Now if you want to do an IOT internet of things which requires which is enabled with this computer vision capabilities then those people who are working for the IOT they have to be a good understanding or a data scientist working for an IOT. These are different job roles for this computer vision role. Now this particular lecture will make you sure that how you can build your good data, how you can prepare the good dataset so that you can build a very good computer vision model. So here you can see that what we mean by good dataset is a dataset which can train the model with a good set of examples is called a good dataset. Now let us see what are the techniques when we say that you want to train the model machine learning models for computer vision what are the techniques which are needed for doing this kind of training before that what are the techniques which are needed to get this computer vision working. So computer vision as you know that deals with all the problems related to the images and the videos and you know that images comprises of pixels that is in the digital forms and the videos comprises of set of several images. So these are the fundamental principles has to be understood in the computer vision. So with the help of so many pixels in the images you have to now decide whether the object in the image is a cat or it is some other animal that is cat or a dog. So how to understand from these particular collection of pixels to get these object understood. So there are a lot of techniques or the fundamentals that can be tackled with the computer vision and few of them include this technique which is called the image classification. So that means image classification means that the object which is built out of several pixel together is able to classify as the cat. So this is called the image classification. Then the next issue is called the image detection. So image detection means that in a particular image where that object is located. This is also sometimes called the localization. So classification that is there is a cat into that image but where in that image that cat locates called localization. Now to get this particular classification and localization done for that you require to build a segment or segment that particular object into the image. This technique is called segmentation. So segmentation is that given an image you have to now segment the image with a different color in contrast to the all other things in the image becomes the background. So that is called the image segmentation. Similarly inside that image which is segmented now you have to do the image classification. Then comes the next issue called pattern detection and object localization that we have already defined. So the object detection and image classification are the two things that we are going to discuss in this particular session in this lecture. So we are going to go into some models they are called predictive models that means they will do a predictions based on this object detection methods. So let us complete that for a single object so far so good but what about if let us say that your image has more than one objects whether it is having a cat it is having a dog and it is having a duck also there are different classes of objects of your interest in your computer vision and all the objects how they are going to be means identified segmented localized and classified all these things together is called the object detection. So if several objects are there so all the objects are to be now identified and this entire technique is called object detection. So object segmentation instance is also shown over here for object detection you have to segment the images as I told you that there are different objects so they are all segmented here in this case. So whether it is the single object or there are multiple objects all this object detection techniques will be able to handle that. So these are the techniques which are very much needed to build a computer vision application. So let us see this what are the architecture for computer vision. So typical end-to-end pipeline for computer vision architecture is that once first you have to get the input data that is nothing in the form of images of those particular objects which need to be classified first. This objects require a pre-processing that means the noise part in the in the images need to be reduced that object into the image or that image need to be scaled and also sometimes requires to change the color. So OpenCV is the software now that is used for doing this pre-processing before this particular object becomes usable for the training of particular computer vision model. Then the next process is called selecting the areas of interest that is to do the object segmentation and the object detection that is this particular technique will be applied once the pre-processing of that particular data set is complete. Then the next part is that how to identify the features of the objects of interest that is from extracting the data about the feature and then based on this learning the inferencing will be performed that is about the prediction or the recognition that is called object recognition and the feature matching. And finally based on this some action has to be taken if it is an IoT device. So the action could be that it will send some information to actuator. For example, if the IoT device is monitoring a particular traffic in the road and it founds that too many cars are there and almost a traffic jam, then the signal at the junctions has to work that means this action has to be now fed to that particular junction. So therefore, these input data which is which is very much needed for this computer vision, the images that you want to train your model, but these images might be coming from lot of different sources. So all these pipeline will be used to build that particular data set in the computer vision. Now let us see how we are going to handle this pre-processing which is very very important. So the input data set to overcome from various issues, noise, scaling and color changing all this is being pre-processed in that raw data. So after pre-processing you will get a data, a prepared data which will be fed for machine learning model training. So machine learning often depends upon the standardization that means you need to pre-process the input image to make sure that they are of same size. So that the machine learning model for training will not find any issues. Similarly there are some other fine kinds of noise while taking that image from the camera. So all that need to be to be dealt before being used in a data set for model training. So if it is not correctly done, the model might learn the noise part and also other features which are not good for image recognition or object detection. And therefore the flawed, the fundamentally it may go wrong and the model may not give the accurate predictions. So therefore pre-processing is very very essential and we have discussed this. Once the data is pre-processed, the next part is about the data scientists. They have to label the data. So what do you mean by the label? So that means if a picture of a cat is presented, so that label is attached and this label will say that this is a image of a cat. So you require a domain expert to see the image and tell. Similarly the data scientist also is required. So how many minimum number of images are needed to train the model? So these are the different things which is called the labeling your data. So for object detection, you have the images that is taken in different way of the same object and if you have an image of a cat or a dog, you need to label that. That means train, you have to tell, label means you have to tell the model that this is the image of a cat. So this label or the tag to the specific area where the dog or a cat is lying is essentially put as the label and this needs to be done very well by. So this particular labeling is very much essential in a supervised learning. That is one part of machine learning which handles this computer vision that is called supervised learning. So these supervised learning model require this data labeling and that is of our interest that is for object detection and the image segmentation will require this aspect to be completed that is called data labeling. Now feature extraction and prediction is the next step in that process of computer vision. So the feature extraction and the prediction part will be performed by the machine learning model. For that you require, you have to finalize which machine learning model you are going to do to apply on for with a data set for computer vision. So these are the parts of the model training. So the model learns about the features which are present to distinguish between the cat and the dog and it will extract the features. So if it is the deep learning, then these features are automatically learned. If it is a machine learning algorithm, then these particular data scientists, they have to now do a feature engineering and identifies what features are important to recognize whether it is classify, whether it is a cat or a dog. And if the features are relevant in the images, then those features are learned through the pattern and later the model uses it to build a sort of rules for itself and these rules are then used to do the predictions when a new data set is given. So this particular part that is called a feature extraction is a part of a machine learning algorithm to do this. Similarly for prediction and recognition that is also a inferencing part after the training the model will do this kind of recognition that is object recognition and a feature engineering. So this part feature extraction is used at during the training process and in the supervised this requires a supervised learning and then the next part which is called the prediction or the recognition, this is called inferencing and this is used to do a predictions. So we are going to see these kind of things. Now let us see what are the different standard models which are now over the period evolved for doing this computer vision. So the field of object detection is has gone to a tremendous transformation overcoming from the earlier days of very low resolution and very low accuracy models using the algorithms. Why because it requires artificial intelligence or the learning based system like which mimics the human beings not the fixed algorithm which can learn all that details of objects. So therefore over the last 20 years a lot of innovations happened and those models are now finalized and decided as the key building blocks for object detection in computer vision. So let us know them, they are called Regional R-CNN, Fast R-CNN, Faster R-CNN, YOLO and SST that is Single Shot Detector. These models are now standardized and now put in the cloud and they are now given as the service that is the software as a service platform they are running these one of these algorithms into these kind of cloud-based services. Let us go one by one very briefly about them. Why because this is a very little background which we intend to prepare so that we will understand this particular lecture further in a great details. So R-CNN is let us say that first model which you can use for object detection which is developed. So R-CNN the full form is Region Based Convolutional Neural Network. So Region Based Convolutional Neural Network uses three different steps. One is that it will scan the image for looking up a possible objects using the algorithm which is called selective search and those kind of regions where there is a possibility of the object. So they will identify and they are called region proposals. Then it will run a convolutional neural network CNN on top of each of these regions of interest and then take this output of each R-CNN and feed it into a support vector machine to classify the region and or a linear regressor to find out the bounding box around that particular object if that object exists. In other words, so the first proposed the regions and then extract the features and then classify those features for those objects. So in essence these object detection uses this image classification and R-CNN is very intuitive and very slow because it has to first identify those regions of interest that is called ROI from the proposal methods from the input image. So if the input image has 3, 4 dogs and cats, so for every object region of interest will be earmarked and then so many number of for example if 5 regions are there, 5 different CNN will run on each region proposal. Now this is a duplicacy that so many convolutional neural network if let us say image has many objects so many number of convolutional neural network will run and then the output of that convolutional neural network require a classifier. So that is it is whether it is a SVM classifier and SVM classifier will classify whether it is a cat or a dog and then where that particular object lies into the image so it also require a regressor. A regressor means it will give the coordinates height and length of that particular object. So that particular object so therefore both the classifier and regressor will be attached after the convolution. So this becomes quite slow to classify this kind of images using R-CNN. So faster R-CNN is an improvement. So you can see here there is only one convolutional neural network and using this convolutional neural network it will scan and identify the regions of interest and for every region you will be having the fully connected layer and this fully connected layer will now do the both the classifier and the regressor. So this becomes a faster because of the fact that so many neural convolutional neural network CNN is not required only one CNN will do that. So therefore this improvement is called the fast R-CNN. So fast R-CNN perform much better in terms of the speed but there was one bottleneck remaining is that the selective search algorithm for generating the region proposal was not efficient and therefore led to the faster R-CNN. So let us see that faster R-CNN used to replace the slow selective search algorithm with the fast neural network. So it introduced the region proposal network. So it has so this selective search algorithm along with that region search together is called region proposal network RPN. So here you can see that after the CNN doing this scan what it will do? It will identify those selective search using region proposal network. So these regional proposals network will contain a region of interest and then with a pooling layer then it will go ahead for a classifier and a regressor, classifier will classify the object and regressor will then bound a box on top of it. So let us see the unique thing about faster R-CNN is region proposal network. So this in the last layer of the initial CNN the sliding window moves around across the feature map and maps it to a lower dimension for each sliding window location. It generates multiple possible regions and each region proposal now consists of an objects and the coordinates. In other words we have to look at each location and consider k different block center around that. So this was the faster R-CNN. Now then there is another model which comes called object detection model called SSD. So SSD stands for single shot detector like faster R-CNN provides enormous speed gains over the faster R-CNN but does so in a remarkably different manner. So we are not going in much details about that but only thing is that we have to tell that these two models will which perform the region proposal and region classification in separate steps and they use region proposal network to generate the regions of interest. Next they are fully connected layers or positive sensitive convolutional layer to classify these regions. So SSD that is single shot detector does the two in a single shot and that is why it is called a single shot detector. It is tremendously predicting the bounding box and the class it processes the image. So given an input image and the set of ground truth labels SSD does the following pass the image through a series of convolutional layers yielding the several set of feature maps at different scales. Now for each locations in each of these feature maps the convolutional filter to evaluate a small set of default bounding box, these default bounding box are essentially equivalent to a faster R-CNN anchor box and for each box simultaneously predict the bounding box offset and the class probability and during the training the match the ground truth box with the predicted box using IOU. So IOU full form is intersection over union. So let us see all these concepts in very standard model and very efficient also called YOLO version 3. YOLO full form is you look once. So you only look once YOLO. So it is popularly known as YOLO is one of the fastest real-time object detection algorithm that is the speed is measured in frames per second FPS. So this is 45 FPS speed in which this real-time object detection takes place in YOLO as compared to R-CNN, faster R-CNN etc. they are quite slow compared to the YOLO. So R-CNN family of algorithm uses regions to localize the objects in the images which means that the model is applied to multiple regions with a high scoring regions of the image are considered the object detected. Instead of selecting some regions YOLO approaches the object detection problem in a different manner it forwards the entire image to a predicting bounding boxes and their probabilities only once through the neural network. So there are many things are now optimized here in YOLO. So let us go ahead and see some of the details of about YOLO version 3 because this particular model which we will be discussing in the computer vision model of doing this computer vision at the edge of an IoT. So YOLO first divides the image into the grid of 13 by 13 and the size of these grid cells varies depending upon the input size and these cells each cell is responsible for predicting the number of boxes in that particular image. So for each bounding box the network also predicts the confidence that the bounding box actually encloses the object and probability of enclosing the objects being a particular class. So all these things are done simultaneously that is in one go. So most of these bounding box are eliminated because of their confidence is low or because they are enclosing the same objects as another bounding box with a very high confidence score and this is called non-maximum suppression. So therefore what you can see over here is that this is an optimization of bounding boxes efficiently done in YOLO and also this particular prediction and that is the classification and the regression both are done in one go. So all things are being merged together and optimized in YOLO becoming a very fast and a very less space consuming model. So it comprises of so many convolutional layer and then you can see that this will have this particular layer of this predictions. Now various models which we have discussed can be compared their performance with a metric. There are several metrics which is now we have to now understand before we go ahead. So I have already mentioned about intersection over the union for bounding the box. This is the measure which is used or which is based on Jaccard index and evaluates the overlap between the bounding box. So intersection over union is given by overlapping area between the predicted bounding box and the ground truth bounding box divided by the area of union between them. So here you can see through the example intersection over union is the area of the overlap. So this is shown over here divided by the area of the union. So with IO intersection of the union, you have to now decide about the overlapping area to be fixed so that the bounding box fits well with the object. The second method about performance is used that is called precision and recall. So precision is the ability of a trained model to identify only the relevant objects that is the percentage of the correct positive predictions and is given by the true positives divided by the true positives and the false positives that is called the precision. Another measure is called the recall for understanding the performance of the model that is the trained model. So recall is the ability of a model to find all the relevant cases that is all the ground truth bounding boxes. So it is the percentage of true positives directed among all the relevant ground truths. So recall is given by this particular formula. So recall is true positives divided by true positive plus false negatives. So here this particular false negatives means the ground truth not detected correctly and false positives means a wrong detection that is detected and false positive, true positive means a correct detection. So this particular background is being built now we have to see about the computer vision services which are available by the cloud providers and they are called software as a service architecture. So computer vision architecture can easily be taken up by the some cloud service that is running a computer vision model in the cloud. So that you can see that these particular model is provided by these different cloud providers using the technology cloud technology called SaaS that is the software as a service. So most of the prominent cloud provider they provide some sort of computer vision APIs or the services that is Azure, Amazon, Google and so on. They offer some variation of these kind of computer vision service. But the basic fundamentals or the models you have to now decide which can fit into these kind of services. So in that architecture all you need to do is to need to have your images available upload your images and then label them. So labeling is vital why because it is to be done with the help of domain experts and the data science. So data science is quite important again even you are choosing the model or the service through the platform that is software as a service. So once you upload your tagged image or a labeled image that is your data set for supervised learning then it will be applied with some machine learning model or a model which we have covered so far and everything that is completely dependent upon now the cloud provider fully managed that particular cloud. So that is what shown over here first you have to prepare your data with the label and that data is to be uploaded to the cloud there that particular data set is presented with one of the algorithms which we have covered for doing this computer vision or training and it will train the model into the cloud and the trained model is now available to make the predictions. So this particular model which is trained can be accessible with the help of service that is with the REST API and this particular model can do the predictions. So therefore software as a service which is being provided as a platform by different cloud providers for doing this computer vision service is very quite similar it depends upon the choice of the client or the developer. So there might be the fundamental difference in the UPIs and APIs which you are calling the service under the hood. So we will be going in more detail discussing about that. So let us see one such service which is called Azure custom vision service for doing the computer vision that is a software as a service platform. So Azure custom vision is a cloud service used to deploy build the computer vision models this is one such example which we are now illustrating here. So computer vision uses the pretty interesting neural network technique and sometimes it is called that transfer learning because it uses a pre-trained model and then you can bring your own new set of images and you can retrain the model that is called transfer learning. So which applies the knowledge gained out of solving one problem in a domain to a different problem which is also in a related situation. This can substantially decrease the time of creating the models from the scratch. So the features which are provided by this Azure custom vision service is to train the computer vision model by simply uploading and labeling the few images. That means you need not have to have a large number of image yet you can do this computer vision fairly well using this particular technique. So you have to build an image classifier using code-free and code-first approach deploy the model in the cloud on-premise or on the edge device. So this is going to be a useful information for doing for enabling the edge with AI. So here in the picture all things are illustrated. Now you can see that this particular Azure custom vision application is running on the cloud as software as a service. Now over the internet the trained model can be either deployed on the edge and once the camera gets the new image this particular trained model will be able to do the predictions. So Azure custom vision let us see that how are you going to deploy on IoT edge device. So you can see this inside of this Azure cloud service. So it is a custom vision service which runs inside the Azure cloud. Now for that it will ask or it will build a image classifier based on one of the algorithms which we have covered earlier. So it will build a custom vision using custom vision and image classifier using some deep learning algorithm or a machine learning algorithm which we have discussed. So it will only provide a framework and once the machine learning algorithm is being decided by the client or the developer then this particular model will automatically use that and perform this computer vision. Now once this particular IoT edge module how that let us see how it will be prepared. So first this image classifier is to be built and then camera module which now will be taking the new images is being identified and these together is prepared as the container and using the docker it will be now deployed into the edge. So this is the edge, this is the IoT edge. So after containerization the camera module will come over here which is built into the cloud. Similarly, the classifier is also now deployed on the edge and this is being enabled with the help of service which is running on the edge that is called Azure IoT edge runtime that we have already covered in the previous lecture. Once these models that is for the image classifier and the camera capture is in place and this image camera capture and the image classifier they communicate with the help of a protocol called HTTP. So HTTP is very close tight net protocol they have to work very closely. Now this particular deployment will enable the edge to do the inside. So this particular for doing more detail inside that particular pre-processed information or a data or insights whatever is being inferenced by IoT edge will be sent back to the IoT hub that is into the Azure cloud. So send the results of an image classifier to IoT hub for making the business decisions. So the topic is to create an image recognition software or a solution with an IoT edge and Azure services. So you can see that lot of applications uses this image recognition and maybe that this particular application such application needs a solution that is vision based solution such as scanning the fruits for quality for segregating into the or classifying into the quality with a medium quality, low quality or high quality and vegetables and also there are various scanning at the airport for different baggages, self-service, checkout there are many more applications. So let us see how that is all can be done with the help of this IoT edge IoT. So in this particular lab session or lecture some of the components some of the hardware which is needed is that Raspberry Pi 3B or better and an USB camera and a speaker. So this is the picture of Raspberry Pi board. So Raspberry Pi board is a standalone computer and it has enough of processing power and it is the device has a limited memory within it. So Raspberry Pi 3B plus board has 1 GB of RAM and now a newer version that is Raspberry Pi 4 has come. Now this particular board will be used to prepare to make this edge, IoT edge. It will work as an IoT edge. And other thing which you require is a desktop with Ubuntu installed in it. And you will be needing since it is a image recognition solution. So you will be requiring a USB camera and that USB camera that is an IoT camera which is connected through a USB port here into this Raspberry Pi and this camera is a device or it is a thing. Now it is this particular IoT, this will become an IoT edge. So this is internet connected and this will be now in turn will be connected to the cloud, will have the connectivity to the can be established. Now there are two more icons which are shown, you require a Ubuntu machine or Linux based machine and then another thing is you have to need a docker. So all this is a requirement for this particular exercise. So let us see that once you have Raspberry Pi board then you have to install the Raspberry Pi operating system and for that you have to now do many things like installation. So the installation of a Raspberry Pi operating system called Raspbian has to happen to run that particular board. Then on the cloud side you have to create an Azure services. So that is called an Azure IoT hub on the cloud platform that is offered as a software as a service. So this Azure IoT hub you have to create your details for that devices. For example, you have an IoT camera and Raspberry Pi. So these devices need to be registered here in Azure IoT hub runtime. So once it is done then you have to install on the Raspberry Pi Azure IoT edge runtime on Raspberry Pi. So you have to download the configurations that is being described and then you have to deploy. Now then again you have to install the command line interface Azure CLI so that it may accept CLI commands from the console and then you have to deploy this IoT edge to the device. So you can see from once it is being prepared this IoT at the IoT hub it will be containerized and that particular services are now deployed on IoT edge device. So this particular deployment of an IoT edge runtime from the Azure IoT hub is done through a technology which is called the Docker. So that is what we have already mentioned. So let us see that once these requirements are done then let us see the solution how it look like. So that solution will scale from Raspberry Pi to the desktop. So let us see that this is an IoT camera and here you have the Raspberry Pi board which is running an Azure IoT hub runtime on the board and so this is an environment which will become like an IoT edge. So this particular IoT edge now will be installed with the using the Docker, the trained machine learning model and this particular camera will take the picture of a fruit and it will be able to detect and display that this is the fruits. So let us see that now again come back to our point to understand more about these setup details. So we have to create the classification model using the Azure custom vision and Azure custom vision services is simple to create an image classification machine learning model without having to be the data scientist or a machine learning expert. So what you have to do now you have to become like a data scientist, you have to upload prepare and upload the labeled images here in the Azure custom vision service and then you can train your model and once the train model, once the training is done then you have to evaluate the model whether the model is properly trained to do the correct predictions or not. So this particular classifier may be further be improved by adding more data set or more samples. So let us see that how you are going to create this particular model. So you have to create the custom vision service and then you have to gather the initial data or the images and then you have to upload and train your model and here there is a train button. So once you train the button then the model will be trained and once the training is done then it will show the performance of performance metric of the train model. So the precision is 97.4 and the recall is 93.3 both are fairly good. Now let us see that computer vision prediction thresholds you can set. So here this is the example that if your model is not having the accuracy and the precision as per your need then you can adjust, you can train the model by modifying this threshold. Let us understand about that. So custom vision offer the fluent prediction threshold adjustment to improve the model performance. So in this case we prefer the higher recall over the higher precision. This is important not to lower those threshold too much as the model performance will suffer significantly and having the low probability threshold will increase the number of false positives. So if the model is supposed to be deployed in a production settings we cannot stop the production line for every false positive detection by the model. Therefore, the key performance indicators or index KPIs will be as follows and will be tunable to improve the model performance. So the main matrix to optimize is the MAP. So this particular average of all that precision and recall is MAP metric. So this particular metric cannot be lower than 85 percent let us say that if we optimize the main metric. So this means that recall and precision are also equally important and they should stay above 80 percent. If that is to be enforced in this particular threshold then your model will be giving a good accuracy and this is what is shown over here. So once these models are trained and trained up to the satisfying performance using this KPIs then this particular classification model classifier is now ready and you can now deploy with the help of Docker. So you have to now containerize this trained model on the cloud and then deploy it on this particular edge. So that is what is the next step. So once it is containerized then you have to download and it will be shipped or deployed to the IoT edge which is nothing but the Raspberry Pi board which is being built or registered with the Azure IoT Hub. So this is some of the more detail that you have to now build the solution and you have to deploy using Docker and container you have to push it on the IoT hub or an edge IoT, IoT edge and then you can also monitor it, its functionality of that IoT edge. So that was a brief of the discussion on the practical side also. So let us summarize this lecture what we have understood in this lecture. We have understood about the computer vision fundamentals that computer vision needs some of the machine learning models. They are called image classifier to build the image classifier. So different machine learning algorithm or the models which we have discussed is R-CNN based models and YOLO based models. Then we have seen about the Azure computer vision service that is a software as a service. We have also told that all other services provided by different cloud vendors are similar in nature. So this particular use case taking from Azure is good enough to understand the functionality of every other solution provided by different vendors such as Amazon and Google. We have also covered the use case and the practical lab session here in this particular lecture. Thank you.