Page 86 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

74 Chapter 3 Application, algorithm, tools directly related to deep learning

Figure 3.9 Convolution layerdthe kernel.

3.2.2 Convolution layerdthe kernel
Convoluting an image with 5 5 1witha3 3 1 kernel to
get a 3 3 1 convolved image is demonstrated in Fig. 3.9.Image
dimensions ¼ 5(height) 5(breadth) 1 (number of channels
(e.g., RGB)). In the aforementioned demonstration, the green
section resembles 5 5 1 input image, I. The element which is
present in carrying out the convolution operation in the initial
part of a convolutional layer is known as the kernel/filter, K.We
have selected K as a 3 3 1 matrix [15].
Kernel/filter, K ¼ 101
011
110
The kernel shifts nine times because of stride length ¼ 1 (non-
strided), every time when performing a matrix multiplication
operation between two elements K and the portion P of the image.
The movement of the kernel is represented in Fig. 3.10.
The filter moves to the right with a certain stride value till it
parses to complete width. Moving on, it hops down to the begin-
ning (left) of the image with the same stride value and repeats the
same process until the entire image is traversed.
Generally, the first convolutional layer is responsible for
capturing all the low-level features such as edges, lines, color,
gradient orientation, and so on. When layers are added, the archi-
tecture adapts to the high-level features as well, giving us a
network which has the moral understanding of images in the
data set [15]. The convolution on an M x N x 3 image with 3 x 3
x3 filter is shown in Fig. 3.11. The convolution operation with
stride length 2 is shown in Fig. 3.12.