Deep Learning) and Convolutional Neural Network are the most trending machine vision subjects in recent times. The artificial intelligence studies (roots at 1950s and completed its theoretical maturation in 1990) lives the golden age with the increasing processing power, the access to big data, and the improvement of imaging equipment today.

The motivation underlying deep learning applications in the field of machine vision is the automatic provision of definitions which are difficult-to-understand with classical image processing methods. In order to do this, instead of describing each object according to its properties, by giving to the system a large number of examples that we already know, it is expected that the system will take out common and different aspects of these examples. So when any picture is shown to the system, it can tell us which one looks the most like before.

Example 1: consider an image processing application that finds the model of each of the brake discs in the image above. With the classical approach, we can make a separation by looking at the number of holes and grooves on the surface. This approach is likely to work.When a new type of disk arrives in the system, it will be needed to see the parser feature of it, then code the feature and generate a new version.

 

object identification with classical methods, by classifying the number of large holes, small holes and grooves

Deep learning is primarily a solution to such problems (rewriting code for each new situation). This is what the word learning here means in general.Once the system has learned the classes, it will try to parse each picture it gets according to these classes. The more images the system is fed the more it will learn, and the healthier it will work. It will be much easier to introduce a new product to the system than in the old one.

In the above example, classifying by classical methods was easy because the picture contains geometric shapes. In real life, we generally come across much more complicated applications that we can not express with geometric shapes (face recognition, morale analysys like sad, happy or mad, and behavioral tests, etc)

 

Example 2: In the above image, let's consider an image processing application that determines sea beach and sky. It will not be possible to talk about geometric patterns here. In this example the color seems to be the decisive point for classifying, but the sea and the sky will be very like to be confused. If the picture had been taken in the evening hours, the beach, sea and the sky would appear in a completely different color, and the shape of the sky would change constantly according to the state of the clouds.The solution of such a problem is really difficult with classical approaches. "Classification" methods are used for such problems. Deep Lerning also commonly uses the CNN classification method. The basic logic is to allow the system to gain insight into examples by giving a large number of examples of the same group. When giving sample images, the relevant regions are labeled (Sea,Beach, Cloud). After the learning phase (train) has been completed, the system looks at a given picture to see what labels it has learned previously. In the above example, a large number of clouds, sea and beach tags are taught to the system. The more systems are fed with different images, the closer the results are to reality.

As in the above image, each region is reported to the system by a label. Offline Training which is extracting the information of the classes (indicated by tags) of the system from all images fed is done. Training is usually a long process. Generally good graphics processors are an advantage because the workload is mostly on the graphics processor. Fortunately, the long running train is done once. The system will decide quickly when it is running.

In summary, in the deep learning process, there are classes and labels. The class can be in the size we want. (In the image above, it is 3-dimensional, for example: Sea cloud and beach.If we wanted to separate objects such as green space, human, ship, etc., we would have to find pictures containing those images and also to teach areas containing those tags.

Detailed study logic and sample projects of Deep learning, CNN and other classifiers will take place in our blog pages.
We as AKU also implement deep learning methods to find the types of errors that can be difficult to model, due to continuous variations in area, size, color and shape. We use Deep Learning in projects like:

• Detection of Fabric Defects (oil trace, weft warp offset, tears etc..its very hard to set up a common application in classical models because of the possibility of different defect types in each fabric type).
• Detection of defects on leather (it is difficult to catch defects with classical methods because it contains natural tissue)
• Detection of broken glass bottle mouths (it is difficult to model it with classical methods because it can be broken any form)

 

Bottle mouth fracture NOT OK set. At least 1 photograph of each type of defect is taught to the system.

 

Our deep learning application that detects fractures on the bottle mouth.

The area of use os Deep Learning is becoming widespread in technological areas such as face recognition in crowded mass (stadium, subway etc.), recognition of surrounding objects in images taken from streets, learning environment for autonomous vehicles.

As AKU, we are developing a deep learning-based machine, especially in textile field, to teach and find any kind of fabric defect. In addition to the methods used by classical systems, we achieve much higher performance by supporting them with texture inspection, background estimation and deep learning techniques.