Deep Transfer Learning for Computer Vision

Hi, everyone. So, today we will be talking about deep transfer learning for Computer Vision, and we'll be covering a couple of real-world applications in the context of the semiconductor manufacturing industry at nanoscale.

A bit about ourselves, so I'm Dipanjan. I'm a data science lead at Applied Materials, also a Google Developer Expert in Machine Learning and an author. Over to you, Sachin.

Thanks, Dipanjan. Hi, everyone. My name is Sachin. I lead the data science competence from Bangalore in applied materials. I have an experience of more than a decade handling multiple enterprise-wide use cases. And in this particular scenario, we'll be discussing about semiconductor industry related AI use cases. So let's jump into it, Dipanjan.

Okay, so today we are going to talk about defect classification. So I think all of us know about Moore's law, where it states like every two years that transistors in a chip gets doubled. And how it is possible, like it is possible by the atomic level engineering, what companies like applied the day in day out they do. And for that, they need to develop these recipes. And these recipes are not easy to develop.

The process engineers who work on these recipes have to deal with lot of defects while refining these recipes. They deal with these high end image capturing techniques, which these tools like laser tech and AFM, which is Atomic Force Microscopic Tech tools, are taken into account. And in these particular cases, some of these techniques are very destructive where the wafers are destroyed. That's where AI can help by figuring out these defects and have a lot of cost savings and time savings.

So I'll give you a quick overview of how we are leveraging AI techniques and defect detection for AI semiconductor manufacturing. So if we talk about the defects, like we deal with multiple type of defects. In this particular use case, we have defects like particle, large particles, small, fit and all. You can see there is a lot of noise and variation. So it was not easy to have one single algorithm or one single model to deal and find the whole classification done. What we have done here is we leverage multiple classifiers, a stacking technique where we input the image. Then we do a lot of image processing to different techniques. We do the denoising and then we feed that image into first level of classifier, which figures out whether there is a defect or not. If there is a defect, then we try to figure out the second level of classifiers, whether it's unknown, or a large, or a small defect.

Even if we find that, then there are different types of image processing techniques we apply to bring that effect out of the background. Then we try to do further analysis, like whether there is noise or the threshold levels are ... The noise levels are below threshold so that we can directly go to the fourth level of classifiers, which are to decide the ultimate class, whether it's a pet or a particle. In this way, we try to leverage transfer learning-based techniques to figure out the classification approach, what we have got.

So, I think Dipanjan will get more deeper into it and we'll ... Over to you Dipanjan. Here, we try to apply a hybrid classification system where we fine-tune a pre-trained ResNet model to extract the deep features. We also use traditional image-processing-based features like gray-level co-operance matrices and zonic moment features. We combine these to form a fusion feature vector, and we pass it through a deep learning net to do the final level of classification. And we also do some image augmentation, especially in cases where we have less data, just like in these cases, because we can't afford to have a lot of data, especially with regard to semiconductor-level electron microscope-level scans. And we also leverage pre-trained models like ResNet-50 in the backend of a object-detection model like a faster RCNN to detect and count the number of defects, because sometimes you also have to give the defect count besides the type of defect. With regard to our evaluation performance, as you can see, the performance is 95 percent on a holistic overview in terms of the four types of classifiers, and we do a per-classifier level performance also to understand the level of performance we are getting at each specific hierarchy in our overall hierarchical classification system.

Now we will discuss about an unsupervised learning case in the context of deep transfer learning. Over to you, Sachin. Thank you, Dipanjan. So, as we saw, like the last case, we had prior knowledge of type of defects, but that's not the case always. Like many of the times we are not aware of how the defect is getting induced in this whole recipe, and there we need to do a lot of analysis, like what is the defect count? What are the morphology of the shape? What are the elements, like whether it's nitrogen, aluminum? Like at different parts or different steps of the recipe, we can encounter different foreign particles which could result into defects.

So because of lack of prior knowledge, we go to the unsupervised technique where we have built an interface where a user can browse the location where we have got all these defect images. Once the images are fed to the system, we use the ResNet-50 model to extract the features. These feature vectors are then fed to algorithms, the various clustering algorithms of what we have got. You know, get the higher level of clustering which is then provided to the user to decide how many clusters they want. And we go to the finer level of clustering like agglomerative clustering or affinity propagation clustering.

Based on these clustering results, we figure out the right clusters as well as the outliers and that is presented to the user. So this makes the whole process very reproducible and repeatable which is otherwise very subjective between different manual users. So over to Dipanya to get more into details of this whole algorithm. Yeah. So again, here we leverage a pre-trained ResNet model along with computer vision specific features like the hierarchic features of the zonic orthogonal projection moment features and we get those fusion vectors which we typically pass to our clustering models. And that is where we use hierarchical clustering in the terms of forming that dendogram, getting that right level of cutoff forming those clusters from the data itself. And then we also isolate the singleton clusters, so which are basically very different from the regular clusters in terms of similar kinds of defects. And this is a view you can see in terms of the interface which can show the different clusters to the user and then they can save it and leverage it in the future to even feed it to a supervised learning problem. And that brings an end to our whirlwind tour of Deep Transfer Learning for computer vision. Thank you.