How Convolutional Neural Network Extract Features?

Convolutional neural network is a deep learning artificial neural network. Researcher design convolutional neural network to solve the complicated challenges that a computer faces to understand image patterns.

Feature extraction procedures of Convolutional neural network

The following figure shows the way in which CNN extract unique features from input data.

1. At low level feature (1^st layer), it detects simply horizontal and vertical lines present in the images.
2. At the 2^nd layer, it identifies the shape of the object because it can detect different corners on input images.
3. At the 3^rd layer, CNN computes more complex feature identification. E.g. extracting differently structured lines and shapes on the surface of objects.
4. At the 4^th higher layer, CNN is more power full to identify each tiny object structure.
5. And while going to a very deep layer it can identify a very small piece of structure. So, using this pattern recognition ability, CNN can identify images unique features and make ready the data to be classified by the classifier layer.

Feature map means the number of unique image features fabricated by convolution process. When we increase the image size, the number of features computed by convolution layer will be increases. For more accurate feature extraction, apply padding on the image size is preferable. Because it convolution visits the pixels of images found at the corner and margin of the image matrix more than once.

Architecture of CNN

CNN has feature learning and classification layers. The feature learning layer is used to extract image features. The classifier layer of CNN receives extracted features from feature learning layer and perform image classification to medicinal plant part.

The feature learning layer of convolutional neural network contains nonlinear activation function and pooling layer. The nonlinear activation function keeps the output of the input layer to be nonlinear for further process.  The Polling layer reduce the feature map of image pattern.