2 min read 26 views

Using a CNN

Using a CNN_image

Convolutional neural networks are perhaps the most impressive image classifier techniques being used today. They manage picture and video information where the task is to learn the "highlights" of the picture for different applications like picture arrangement and semantic division. Let's say that we are making a framework that classifies various types of birds. We will walk through the steps of using a CNN.

•  Collection of information: 
The principal task is to gather good data. The dataset can be in picture or video form and we need to regularly gather in excess of 1,000 pictures for each class. We need to gather around 1,000 pictures for every one of the class, and the more pictures we have, the better model will be able to generalize. 

•   Defining the network architecure: 
In a convolutional neural network, we can define the number of layers, which types of layers, and several hyperparameters such as convolutional kernel size, number of filters, etc. We also want to standardize the dataset so that all images are the same size.

•   Training the network: 
At this point, the network is prepared with every one of the pictures set to an equivalent standard size. "Inconsistent" size of pictures cannot be prepared. The network takes input pictures in "groups". The network is prepared for a particular number of passes or "epochs" and a "loss" is determined toward the finish of every epoch, the loss characterizes the boundary esteems to upgrade in a neural network model and the lower it is the better the model. At first, the loss will be high yet then the neural network changes itself to bring down the loss through a method known as "backpropagation". The lower layers catch low-level highlights, for example, edges or bends however, the higher layers catch undeniable level highlights like eyes, ears, nose and the face structures. The idea is that low-level highlights join to frame more significant level highlights and in this manner CNN defines its outcomes dependent on the relationship of seeing individual highlights, for example, sparrows have little paws and hawks have enormous hooks additionally sparrows have more modest snouts while birds have bigger bills besides sparrows have little eyes while falcons have huge eyes. 

•   Testing the network: 
When  the network is tried on concealed pictures and the outcomes are analyzed if the model is precisely making expectations that implies we have effectively prepared with the correct boundaries yet assuming the model is inadequate with regards to exactness, we may have to explore different avenues regarding the boundaries like the quantity of layers or the learning rate or possibly add more information.