Figure 3.

(a) Schematic of the base CNN model showing the flow of data along with the shape transformations (in height × width × channels format, where applicable) it undergoes at each layer of the model. (b) The spectrogram pre-conditioning layer, comprising one-dimensional Laplacian of Gaussian (LoG) operators applied at two scales, trainable convolution layers (Conv) and a concatenation layer (node C). (c) The quasi-DenseNet subnetwork, comprising four quasi-dense blocks. (d) Illustration of the difference between a typical dense block (as in DenseNet) and a quasi-dense block. FCN: fully connected network.