How to implement image processing with deep neural networks?

We’ll have a comprehensive dem over will implement image processing with deep neural networks so let’s talk about the very first topic now this is a big challenge in our industry currently why do I say that very simple example have you guys heard about NLP I think yes I have shown you some implemented this yes also you guys have heard about computer vision all right so what basically these topics are is or we can also call voice processing now and what are these topics is you have got a pre-tpre-trained that means if you are talking about NLP you have got.


Let us say if you are talking about with within you are talking about some kind of sentiment analysis what customer is talking about is he talking positive negative good bad about us if you want to huge if you want to do this kind of processing on in bulk you can use NLP sentiment analysis here now to do this what you need is you need a very good corpus file a back-end file of it yeah so why do I say we need a good back corpus file is we need to have all the possible combinations here itself so that I can train my neural network all right these files are usually very heavy if I show you guys an example of red if you look at this is a corpus file available with me for flowers.


If you have if I want to if you give me an image and if I want to classify that image in one of these flowers I need this kind of images as a backup to train my model saying that this is a tulip here for example or this could be a jeweler so there could be some positive examples that could be some negative examples we don’t know about so this is the first challenge that we usually face how to get this kind of corpora so if you talk about voice processing I’m not sure if you guys have I shown you anytime anything on to voice ever no right lastly you said something and it perfectly pipe right yeah perfect so that voice processing imagines how many levels of voices it had to record to understand so that corpus that we had was from Google, for now, it is free.


We are using it Google was kind enough to gather certain people from certain ethnicity certain regions to speak certain type of words so that whenever I speak something he tries to match my accent and my word with the nearest value and tries to predict that I say that if you observed last week whatever I say it not hundred people was a match but yes majorly it understood what I was trying to say so the complete challenge for any neural network here lies to gather this kind of combinations so that we can train a network that is what we call it as augmentation okay you can relate that topic with augmentation now what do I mean by augmentation let us say you have given this as an image to your neural network saying that this is a cat now what if I tilt my image.


What if I rotate my image or what if I crop my image what if i zoom in my image whatever you do with this if in case if you train a neural net up onto this only one picture saying this is cat and if you present this to him a simple neural network it will not take it up why because if you observe the pixels present here and if you take the same locations in your image current image there might be empty space so obviously it’s more not going to match it so it’s very important for us to make sure that before we give a data to a neural network we have done a proper augmentation augmentation means we have done this kind of changes now it is in our hand whether we have to do this so there are certain functions available within neural networks which will allow us to do augmentation so it will ask us if you want to rotate if you want to flip if you want to crop if you want to zoom whatever you want to do all right sometimes even it allows us to change the colors also so this is called data augmentation.


If I just present this it’s not a major topic but yes it is much talked about because not all of us have very good access to all these things all right so please remember it says we may not have a big data set to create more data so how to create a more data you can create more data using your augmentation technique like this all right so just take an example of an augmentation what we can do is this is an image which belongs to a dog if you want to have what you say transform the image let us say from color he has transformed into a grayscale from there I am giving it to a CNN now, in this case, it is easy for us to identify what if tomorrow the snow and stuff is not there only this much face is available we are not sure how comfortable your neural network will be adding somebody’s cameras on this yeah.


How comfortable the neural network will be to identify that it’s a husky right so in those cases augmentation will be of very good help so this is one case where you can say I need augmentation and definitely in computer vision you need a good backup of data otherwise neural network is of no use okay what all we can do inside of my patience oh you can do flips you can do rotations cropping scaling color jitter means changing the U and all this part of it other creative techniques could be if we talk about convolutional nets it could be translation rotation stretching sharing lens distortion.


You can do a lot of things now you might be wondering what are these terms this will come in the first week of computer vision where we will use certain filters on top of an image such a way that we will change the look and feel of the image so this is something like you can we can say on Instagram you know those instant filters are there yeah some of the filters you know they change the background they blur the background and so this is what the filters are so you just have to define a filter which could be multiplied with an image and you will get a new image out of it that’s it so this is what is by lens distortion shear all right so this is augmentation any questions pretty straightforward.


In computer vision we’ll see the real application of it for the part of our neural network of it we usually have a good backup data it should not be big trouble all right now coming to weights initialization this was common I think in the first session a lot of people asked do we actually put randomized weights I will say it by default we do it but if we don’t want to do it there are certain different type of weights available with us let’s see what are those weights so one yeah so this is the total weights that we can say we have it in in our hands either we can use zero initialization.


We can put all the weights as zero yeah we can put random my random initialization that is what we do currently we are something for Xavier initialization we have something called chi and there are many more so if if you are a researcher in a company and all if you want to do it you can even develop your own weight scent right now in today’s case study that I’m going to show you we will see random versus a chi so first wheels we will start with a chi will implement a model and then we will change it to random and we will check how the accuracy changes alright so these are the two common ones if you want to use it so any idea where I can use 0 initialization because you remember.

Leave a Comment