Welcome to the exciting world of deep learning! Now, together with you, we will meet a miraculous architecture like CNN in the name of image classification.
CNN is such a magical structure that, believe me, you will lean back in your seat without any effort and leave the job to a connoisseur. In addition to computerized vision, it also contains deep learning, bringing with it many capabilities such as object recognition and classification. Therefore, the use of CNN in deep learning is inevitable!
I also want to share with you the valuable information I have received thanks to Andrew NG. During my Master’s education, I had the opportunity to use CNN on many projects. In addition, we have many different neural networks, of course! But if the images are available and the classification is to be done, CNN won’t be found.
What are these Convolutional Neural Networks?
As the name suggests, convolutional neural networks apply a convolution layer to each pixel for images contained in the dataset. After briefly mentioning it, let’s find out where the foundations of evolutionary neural networks are based. Its main structure contains the foundations of computer vision and deep learning. Even computerized vision is an area that is more active in our lives with deep learning.
CNN neural networks achieve a strata-different output by moving the filter you specify step by step on the image. If you want to study these project-based steps, you can study classification writing with Keras. You can also find the creation of the required layers in the article as Python code when creating the CNN architecture.
Instead of the classic ConvNet architecture, you can also use different architectures. Different models of the CNN network are included in articles and projects in the literature. Based on these results, it is possible to achieve more accurate results by using different deep learning models in image classifications.
- Microsoft RestNet
As an example, if we talk about a data set that is very often used in object recognition and classification, there will undoubtedly be by MNIST data. If you’re interested in deep learning, you’ve definitely come across figure recognition videos or studies. At this stage, before training the machine, it would be useful to know how the neural network works in the background.
Digit Recognition with MNIST Data
I had a chance to apply many different filters during the process of processing the image. Laplace, Sobel, Gamma and many others.. A kernel is defined for these filters, and then hover over the image to be applied. Especially if we consider this stage, we can understand how the CNN structure works. First of all, stage-by-stage navigation is provided with the block structure.
If you need to examine MNIST data, it is handwritten adapted data. Take, for example, the digit 9. Let’s say I showed you the number 9 on a board. You can immediately tell that the number I showed you is 9, can’t you? However, computers may not be able to detect an image that they have not been taught as quickly as we have. Then they estimate based on the neighborhood values and density of these pixels. In addition, evolutionary neural networks also process pixel values when detecting edges.
As an example, let’s look at the zoomed state of the number 9 as if it were a machine.
In this handwritten image, it is possible to estimate the densities of the White and black pixels contained in 9.
What I wanted to say was how the image structure was analyzed. In the meantime, it should be noted that CNN will work in two-dimensional images (black and white), as well as in three-dimensional images where the color channel exists. I’m leaving a source link that I think will be very useful for understanding RGB and Grayscale structures.
Let’s analyze the numbers in the RGB image above. The digit 6 is represented by width, the other digit 6 is represented by height, while 3 shows the number of channels. Blocks indicated by a yellow box are the effective convolution layer on the image. To better understand evolutionary formation, you can follow the Convolutional Neural Networks course on the Coursera platform described by Andrew NG.
🚀 Let’s provide our data through Kaggle or any database source and train the CNN neural network and perform object recognition together! It’s not too late for anything! Note that a high accuracy rate (with exceptions) is usually desired when setting up a forecasting model. Of course, this situation varies according to different data, projects, and goals. How about coding the CNN neural network in another post? I wish you healthy days.
- Coursera, Convolutional Neural Networks, Andrew NG, Younes Bensouda Mourri, Kian Katanforoosh.
- Gaziosmanpasa, JOURNAL OF SCIENTIFIC RESEARCH (GBAD), Deep Learning models used in Deep Learning And Image Analysis, Ozkan Inik, Erkan Ulker.