Proposed architecture

Figure 12.10shows the proposed architecture of the four-layer deep CNN. It is used to classify the entry of non-humans in the intrusion detection system (P Arokianathan et al. 2017) and the presence of the disease in the leaves of the paddy crops. The architecture includes four layers of CNN followed by batch normalization with a ReLu activation function. The normalization is included to stabilize, to increase the speed and finally to increase the performance of the network. Each CNN layer includes a kernel size of 3, 3. A dropout of 0.25 is included in the second, third and fourth CNN layer to reduce the over fitting of the features (Sivakumar and Rajalakshmi 2017) across the network. The first and second CNN layers contain a filter size of 32, while the third and fourth layers include filters of sizes 64 and 128, respectively. A max pooling layer is included in the second and fourth layers of CNN. This pooling layer is used to reduce the dimensionality of the input feature and thereby increase the computation speed.

The four layers of CNN act as a feature extraction phase of the network. The second phase of the network includes three dense layers used for the classification of the images based on the feature extracted. The three dense layers consist, respectively, of 512, 128 and 4 as their output units. The first and second dense layers consists of ReLu activation function, batch normalization and a dropout of 0.5. The last dense layer consists of softmax activation function, loss as categorical cross-entropy and optimizer as Adam. The softmax function is used as output in the neural network to get “k” different possible outcomes. The categorical cross-entropy is a loss function applicable to an instance that belongs to one class only. Adam is a combination of Stochastic Gradient Descent (SGD) with momentum and RMSprop.