Understanding Image Classification Using TensorFlow Deep Learning - Convolution Neural Network

1.1 Neural Network

In machine learning algorithms based on the configuration and operation of biological neural networks. This concept appeared to simulate the process taken by Warren McCulloch and Walter Pitts in 1943. A neural network consists of a single unit called a neuron. As shown, the neurons are located in a series of layers (Figure 1). Each layer of neurons is connected to the next layer of neurons. For these compounds, Data comes from the input to the output layer. Perform simple mathematical calculations for each individual node. Then transfer the data to all connected nodes.

Figure 1.

Neural network

The last layer of neural networks is related to increase computing power and experience. This brings deep learning, makes the neural networks technical structure more complex, and solves various tasks that previously could not be solved effectively. Image classification is a key instance.

1.2 Convolutional Neural Networks and Image Classification

The CNN is a special structure of the ANN proposed by Yann LeCun (1988). CNN uses several functions of the visual cortex. One of the most common use of this architecture is classification of image. For example, Facebook uses CNN as its automatic tagging algorithm, Amazon for generate product recommendations, and Google to search for user photos.

The first layer is Convolution layer. Input Image (matrix with pixel values) (figure 2). Imagine reading the input matrix from the top left corner for pic. The software then selects a small matrix called a filter (or neuron or core). The filter then generates a convolution i.e. it moves along the input image. The task of the filter is to multiply that value by the original pixel value. All these multiplications are additive. Finally, get the number. The filter reads only the image in the upper left corner, so move the unit further to the right and perform the same operation. After passing through all filters, the matrix is acquired, but less than the input matrix.

Figure 2.

Input layer of matrix indicate the first hidden layer

From a human point of view, this process is like recognizing the borders and simple colors of a picture. However, the entire network is required to identify high-level attributes such as trunks and large ears. This network consists of several convolutional networks that combine nonlinear and pooling layers. When the image pass over the convolution layer, the result of the first layer enters the input to the second layer. This also further for each convolutional layer.

A non-linear layer is added after each convolution operation. There is an activation function that brings nonlinear characteristics. Without this attribute, the network would not be powerful and the response would not be variable cannot be models.

The nonlinear layer follows by the pool layer. Applies to the width and height of the image and performs a down sampling operation on them. As a result, the volume of the image is reduced. In other words, if some functions (as for example boundaries) are identified in the previous convolution operation, the detailed image of the subsequent processing becomes unnecessary and the image is compressed to less detail image.

After completing a series of convolutional layers, nonlinear layers, and pooling layers, you need to connect the fully connected layers. This layer gets the output information from the convolutional network. Attaching a fully connected layer to the end of the network produces an N-dimensional vector. Where N is the number of classes from which the model selects the desired class. This piece of code written in python. 1.2.1 Process

The beginning of this section describes the process of monitoring machine learning that is used as the basis for the model.