14.   NEURALNET FILTER PARAMETER ADVISOR

14.1   Parameter Advisor
14.2   License Plate Recognition

---

14.   NeuralNet Filter PARAMETER ADVISOR

NeuralNet Filter has more parameters than the BioFilter and Neural Filter. In this chapter, we will study the Parameter Advisor, which will help you to select values for the parameters. We will use a Car-License-Plate Identification Problem as an example.

14.1   Parameter Advisor
 

Figure 14.1 Parameter Advisor.

The data used in this section is in the �license1\� directory. The Parameter Advisor helps you to choose a set of important parameters for the NeuralNet Filter. The Parameter Advisor will make recommendations for:

• Segment Cut
• Blurring
• Sensitivity
• Internal Weight Cut

You will still need to set the other parameters, but they are relatively easy to set. To use the Parameter Advisor, you must:

• Put all similar images in a directory;
• Set the parameters very high to make sure that the ImageFinder will retrieve the matching images.

We will start with the following parameters:

Segment Cut = 0
Blurring = 50
Internal Cut = 95
Sensitivity = 95

The reason is: the Advisor will only look at the parameter-range up to a certain point. For example, if you set the Blurring = 20, then the Advisor will only look at the range of Blurring from 0 to 20; if you set the Blurring = 50, then the Advisor will only look at the range of Blurring from 0 to 50;

The data used in this example is given in the following figure:
 

Figure 14.2 All images in the first class.

The following example will show you how to use the Parameter Advisor. You can duplicate the set up of this example by several clicks:

• Click the �Example/NeuralNet/Parameter Advisor� to open the parameters for this example;
• Click �Batch/Load� to load the parameters;
• Click �NeuralNet/Parameter Advisor� to open the Parameter Advisor;
• In the Parameter Advisor, click �Get Parameter� button; this will initiate the NeuralNet Filter to train and search;
• Close the ImageFinder output. The recommendation will be in the text area of the Parameter Advisor;
• Close the Parameter Advisor.

In the following, we will show you the details of running this example.

The first two clicks load the parameters; alternatively, you can do this manually. The data used is in the �.\license1\� directory. The steps to do it manually are:

1. Data directory is �license1\�. See these images with Windows Explorer.

761SBY~7.JPG
761SBY~2.JPG
761SBY~3.JPG
761SBY~4.JPG
761SBY~6.JPG
761SBY~1.JPG
761SBY~1.JPG

2. Set image, license1\761SBY~1.JPG, as a sample image (Click Key Segment button).

3. Set directory, license1\, as the search directory (Click Search Dir button).

4. Threshold: Average Filter.

Threshold Filter Parameter:
Red: Ignore
Green: Ignore
Blue:  0 � 100; Light Background.

5. Reduction Filter: Real Max

Reduction Filter Parameter
Border Cut = 2

6. NeuralNet Filter Parameter:

Symmetry: Rotation
Segment Size: �S Segmant� (Small)
R(rotation) Type: Type 1
Sample Segment: 50 50 200 200

Set the following parameter to as large as possible to make sure that the ImageFinder will retrieve the matching images:

Segment Cut = 0
Blurring = 90
Internal Cut = 90
Sensitivity = 90

You can do all of the above by:

• Click the �Example/NeuralNet/Parameter Advisor�;
• Click the �Batch/Load�.

To get a recommendation, now:

• Click the �NeuralNet/Parameter Advisor� to open the Parameter Advisor;
• Open the Parameter Advisor, click the �Get Parameter� button; this will initiate the NeuralNet Filter to train and search;

you will get:

761SBY~4.JPG  428892
761SBY~2.JPG  59250
761SBY~3.JPG  535000
761SBY~1.JPG  128000000
761SBY~6.JPG  518000
761SBY~7.JPG  14000000

The recommendation is:

Segment Cut: Very High
Blurring: 23 28
Internal Weight Cut: 16 to 21
Sensitivity: 9 to 14

Remember the recommendation is based on the average of the minimum requirement. The recommendation is generally on the lower side. In this example, we are trying to use one set of the parameters which will fit all of the license plate images, so we will choose the parameter a bit higher than the recommendation.

After a few rounds of testing, this is the set of parameters:

Reduction: Real Max
Blurring: 20
Sensitivity: 25
Internal Cut: 22

14.2   License Plate Recognition

The data used in this section is in the �.\license� directory. Use the Windows Explorer to see the images. This example has 50 images from 9 different car plates. The following are some of the examples:
 

 

   ...

Figure 14.3 The first image in each class.
 

The operation has two steps:

1. Click �Example/NeuralNet/License Plates� to get the batch code;
2. Click �Batch/Run� Button.

47 out of 50 are identified. There are three errors, 3345~1.jpg, 3349~5.jpg and 3349~9.jpg, which gives

Identification Rate = 94% = 3/50.
Error Rate = 6%.
 
 

Return