Beginning Anomaly Detection Using

filter, or kernel. 

The filter goes over the first 2x2 region in the image and sums the element-wise 

multiplication of the values in the filter and the values in the 2x2 region of the image. 

This value is the first element of the feature map, which is a 4x4 layer image. Given an 

nxn filter and mxm image, your feature map dimensions will be an m-n+1 x m-n+1 

dimensional image. In this case, your image is 5x5 and the kernel is 2x2, so the feature 

map is 5 – 2 + 1 = 4x4 pixels.

The filter goes through each region in the image pixel by pixel, as shown in  

Figure 

3- 36

.

Figure 3-35.  An example of one multiplication of the 2x2 filter on a 2x2 section 

of the input image. The filter weights are applied element-wise and produce an 

output value that is part of the feature map–the output of this convolutional layer

Chapter 3   IntroduCtIon to deep LearnIng

---

97

The filter continues doing this until it reaches the right side of the image. After  

that, the filter goes one down and starts again from the left side of the image, like in  

Figure 

3- 37

.

Figure 3-36.  After the operation in Figure 

3-35

, the filter moves to the next set of 

data to multiply over, producing the second value in the feature map

Chapter 3   IntroduCtIon to deep LearnIng

---

98

From here, the filter continues moving right in a pixel by pixel fashion  

(see Figure 

 

3- 38

).

Figure 3-37.  Showing what happens after the filter reaches the right-most side of 

the image. It moves down one (in this case, at least; you can specify how much you 

want the filter to move as a parameter when calling this layer) and then continues 

its operations as usual

Chapter 3   IntroduCtIon to deep LearnIng

---

99

Once it reaches the end, it goes back to the first column and down one row and 

continues its operations until it reaches the bottom right region (see Figure 

3-39

).

Figure 3-38.  The filter continues moving as normal, adding more values to the 

feature map

Chapter 3   IntroduCtIon to deep LearnIng

---

100

The feature map doesn’t make much sense due to the randomness of the weights.

After the two convolutional layers, you run into the MaxPooling2D layer.