Künstliche Intelligenz und Maschinelles Lernen in der Praxis

1. 00:00Welcome to Unit Introduction Computer Vision.
2. 00:04In this unit we will present a short thematic Introducing Computer Vision,
3. 00:09before we get into the practical Take a look at the example.
4. 00:13Computer Vision is the term used to describe that field in machine learning, where you can take pictures on a computer.
5. 00:18the most varied types and How processed and analyzed.
6. 00:23In the Computer Vision area, some tasks, such as face recognition or
7. 00:28the recognition of Gestures or poses in pictures.
8. 00:30Of course, the Recognition of objects in images is important.
9. 00:35It is important to distinguish between: if you only have one object in each
10. 00:40image or several.
11. 00:43It is also important to to distinguish whether you only want to classify an image
12. 00:47or the objects localize or segment.
13. 00:52However, we want to first a step back.
14. 00:55What are the Images for a computer?
15. 00:59On screens, images are naturally colored , but pictures are taken from a
16. 01:04Computer only as numbers and saved.
17. 01:07In this picture, a black-and-white image is by Abraham Lincoln.
18. 01:15In a black-and-white color space with eight bits per Color pixel stands for black and 255 for white.
19. 01:24Of course, the least Pictures black and white.
20. 01:27Most of the time, however, there are for colorful Pictures three so-called channels.
21. 01:31That means for every pixel three values for blue, green and red.
22. 01:37But how exactly can one get out of this Matrix of numbers now recognize patterns or even faces?
23. 01:45Very prominent in this field are convolutional neural networks.
24. 01:50Convolutional Neural Networks are very well known before all things through image processing
25. 01:54or the processing of audio data.
26. 01:58A convolutional neural network has always three main components,
27. 02:03which are installed in a row. The convolution layer, the pooling layer
28. 02:08and the Fully Connected Layer.
29. 02:12You can create multiple convolutional and pooling layers
30. 02:15in a row.
31. 02:16However, before you start using Fully Connected Layer at the end of the network, you have to
32. 02:22Flatten or to German flatten.
33. 02:26We will each be briefly on the different types of layers.
34. 02:31Convolutional Layer use kernels,
35. 02:33different sizes to relevant patterns for prediction.
36. 02:40As mentioned briefly, we have of course not just a so-called channel,
37. 02:43It could be multiple.
38. 02:45These can also be used,
39. 02:47to achieve a common result.
40. 02:51iterate convolutional layers
41. 02:54in a certain order over the image.
42. 02:56And only the result of the kernel calculations is passed to the next layer.
43. 03:02Pooling layers are used for
44. 03:04reduce complexity and reduce complexity pass on data to the following levels.
45. 03:11An example of a Pooling function is the max function.
46. 03:14That means we only take the maximum value of an image part
47. 03:19and pass it to the next layer.
48. 03:23Before taking the results in a Fully Connected Layer,
49. 03:27The results are flattened.
50. 03:29So we don't two-dimensional images or matrices,
51. 03:35because they can do Fully Connected Layer do not process.
52. 03:39Fully connected layers are also classic often part of other neural networks.
53. 03:44In this layer, each node is connected to each Nodes connected in the previous level.
54. 03:52At the end of the course we still have an output layer, for example in three
55. 03:57different classes, dog, cat and mouse.
56. 04:02Even if the subject is not really fits into the rubric "Basics",
57. 04:06so we wanted to briefly.
58. 04:10Under an Adversarial Attack, too German enemy Attack, understood in the context of machine learning
59. 04:16the use of Adversarial Examples, examples hostile to German,
60. 04:21for manipulating classification results.
61. 04:25An Adversarial Example is a specially manipulated input signal
62. 04:29a neural network, which deliberately misclassifications.
63. 04:36These types of attacks are not to Computer Vision.
64. 04:39But here, the examples are very illustrative.
65. 04:43In the left example, it is on the one hand, by means of a fully sprayed stop sign,
66. 04:49that stop signs no longer as such.
67. 04:52In this case the sign was called 50s Shield detected. The right stop sign became conscious
68. 04:59with four small adhesives which had the same effect.
69. 05:04The adhesives were deliberately applied at these locations because the people knew which parts
70. 05:08of the image, the model is particularly respected.
71. 05:12On the right is a glued to see glasses with which you can almost
72. 05:16facial recognition system. man is then no longer perceived as a person.
73. 05:23Students have the same of another university.
74. 05:26For example, if you apply the following image: so you won't be using the Computerv Vision model
75. 05:31as a person.
76. 05:34After this short excursion to the topic Adversarial Attacks here a short note.
77. 05:41If you are more interested in the details, there is a special openHPI course.
78. 05:47Here is the topic of practical introduction to Deep Learning for Computer Vision.
79. 05:53Since this course mainly focuses on Computer vision,
80. 05:56of course, topics will be discussed significantly deeper.
81. 06:00After this short introduction in the topic Computer Vision,
82. 06:03Will we be in the next unit with the practical project.
83. 06:08This is about recognition sign language in pictures.
84. 06:11In doing so, we will present the Convolutional Neural Network.

To enable the transcript, please select a language in the video player settings menu.