Page 271 - AI Ver 1.0 Class 10
P. 271

Rectified Linear Unit (ReLU)

                 This layer is the next after the convolutional layer. It takes the features maps of the convolutional layer and
                 generates the activation map by discarding all the negative numbers of the feature maps.It means all positive
                 numbers will go as it to the system but all negative numbers will go as zero which makes the feature map appear
                 as Non-linear graph with all positive values. This enhances the activation layer and provides better features to
                 the input image for further processing by the next layer in the CNN.

                 Pooling Layer
                 This layer reduces the dimensions of the input image while still retaining the important features. This will help
                 in making the input image more resistant to small transformations, distortions and translations. All this is done
                 to reduce the number of parameters and computation in the network thus making it more manageable and
                 improving the efficiency of the whole system.

                 For example if an image of an animal is given as an input to the CNN  then by just retaining the shape of the
                 eyes, ears and face it is easy to identify an animal. Keeping all the features will just increase the processing time
                 of the image and the complexity of the code for calculating the maximum probability.

                 There are two types of pooling:
                    •    Max Pooling: Max Pooling is the most commonly used method that selects the maximum value of the
                    current image view and helps preserve the maximum detected features.

                    •    Average Pooling:  Average  Pooling  finds  out  the  average  value  of  the  current  image  view  and  thus
                    downsamples the feature map.


                 Fully Connected Layer
                 This is the last and the final layer of the Convolutional Neural Network. After the features of the input image
                 are extracted by the convolution layers and downsampled by the pooling layers,their output is a 3-dimensional
                 matrix which is flattened into a vector of values. These values of the single vector represent a specific feature of a
                 specific label and are redirected to fully connected layers to predict the final outputs of the network. This helps in
                 classifying an image into a specific label based on the probability of the input being in a specific class. For example
                 the input image with a beak belongs to the category of a bird.


                                                                                                        bird     p bird


                                                                                                       sunset   p sunset

                                                                                                        dog      p dog

                                                                                      ...               cat      p cat

                      convolution + nonlinearity     max pooling             vec                       ...


                                 convolution + pooling layers                fully connected layers  Nx binary classification

                                                    Image Classification Using CNN







                                                                                          Computer Vision   269
   266   267   268   269   270   271   272   273   274   275   276