WEEK 3 : Malaria Parasite Detection

Hi everyone! This week, we will give information about the architecture of convolutional neural networks which we will use in our project and the models that use this architecture. You can find second week here.

What is Convolutional Neural Network?

In deep learning, a convolutional neural network (CNN, or ConvNet) is a class of deep neural networks, most commonly applied to analyzing visual imagery. This method is often used in image and video recognition, suggestion systems image classification, medical image analysis and natural language processing.

Cnn uses unique features that make a plane an airplane or snake a snake to distinguish the images given. In fact, this process is unconscious in our brains. For example, when we look at an airplane picture, we can define the aircraft by separating the characteristics of two wings, engines and windows. Cnn does the same thing, but before they detect lower-level features such as curves and edges, and form them to more abstract concepts.

The following are layers and properties of the Cnn architecture:

Convolutional Layer — Used to determine properties
Non-Linearity Layer — Introduction to non-linearity
Pooling (Downsampling) Layer — Reduces the number of weights and controls suitability
Flattening Layer — Prepares data for the Classic Neural Network
Fully-Connected Layer — Standard Neural Network used in classification

LeNet-5

This is the first successful and 7 level cnn model. LeNet-5, which classifies the numbers, was applied to recognize 32x32 pixel grayscale handwriting digits. Unlike other models that will be developed later in this model, average pooling is performed instead of max-pooling in size reduction steps. In addition, sigmoid and hyperbolic tangent are used as activation functions. Unlike other models that will be developed later in this model, average pooling is performed instead of max-pooling in size reduction steps. In addition, sigmoid and hyperbolic tangent are used as activation functions. This architecture has about 60,000 parameters.

AlexNet

Basically it is very similar to the LeNet model because of the successive convolution and pooling layers. But deeper, there are more layers per layer. With 60 Million parameters, AlexNet has 8 layers — 5 convolutional and 3 fully-connected. ReLU (Rectified Linear Unit) is used as activation function and max-pooling is used in pooling layers.

VGG-16

The vgg16 model that is a simple network model and the most important difference from the another models is the use of convolution coats with two or three. In the full-link (FC) layer, 7x7x512 = 4096 is converted into an attribute vector with a neuron. At the output of two FC layers, the softmax performance of the 1000 class is calculated. Approximately 138 million parameters are calculated. As in the other models, the height and width dimensions of the matrices decreases while the depth value (number of channels) increases.The filters with different weights are calculated at the output of each convolution layer of the model, and as the number of layers increases, the attributes formed in the filters represent the lar depths of the image.

See you next week…