Fig. 3
From: On the utilization of deep and ensemble learning to detect milk adulteration
The proposed Convolutional Neural Network for multiclass classification of whole infrared spectra. The architecture consists of one convolutional layer that learns 32 filters of kernel size 5, which is capable of recognizing features directly from the raw infrared spectra. The output of the convolutional layer is concatenated then passed as input to a dense (fully-connected) layer, consisting of 1024 neurons. BatchNormalization, LeakyReLU and Dropout operations are performed in both convolutional and dense layers. Finally, the output layer of size 6 (the number of classes in multiclass problem) is activated by the Softmax function