Table 1.

Structure of the convolutional neural network employed in this work.

Layer	Type	Abstraction-Level Feature
0	2-D convolutional layer with 11 × 11 kernel, padding size 6, stride length size 4	–
1	Rectified linear activation function layer	–
2	Max pooling layer, downsample factor 2, stride length 2	–
3	2-D convolutional layer with 7 × 7 kernel, padding size 3, stride length size 4	–
4	Rectified linear activation function layer	–
5	Max pooling layer, downsample factor 2, stride length 2	–
6	Fully connected layer, 256 neurons	Edge magnitudes: Grayscaled, original image filtered by Laplacian of a Gaussian, result downsampled by a factor of 8
7	Rectified linear activation function layer	–
8	Fully connected layer, 256 neurons	Shape features: area, perimeter, Hu’s moments (Hu, 1962), Zernike moments (Zhenjiang, 2000)
9	Rectified linear activation function layer	–
10	Fully connected layer, 3 neurons	–
10	Softmax classification layer	–

Type: the type of layer. Abstraction-Level Feature: The additional features injected in the layer, if any. Features are injected by concatenation with the input from the previous layer.