Image Knowledge Builder v2

Image Knowledge Builder v2
Image Knowledge
Builder v2
Image learning and recognition
based on the NeuroMem network
Preliminary
Version 2.0.0
Revised 06/08/2016
Image Knowledge Builder is a product of General Vision, Inc. (GV)
This manual is copyrighted and published by GV. All rights reserved. No parts of this work may be reproduced in
any form or by any means - graphic, electronic, or mechanical, including photocopying, recording, taping, or
information storage and retrieval systems - without the written permission of GV.
For information about ownership, copyrights, warranties and liabilities, refer to the document Standard Terms And
Conditions Of Sale or contact us at www.general-vision.com.
HU
General Vision
U
2
1
Table of content
1
TABLE OF CONTENT ......................................................................................................................................3
2
INTRODUCTION ...............................................................................................................................................4
2.1
3
WHAT CAN I DO WITH IMAGE KNOWLEDGE BUILDER? ...................................................................................4
INSTALLATION.................................................................................................................................................4
3.1
3.2
3.3
4
PRELIMINARIES ..............................................................................................................................................4
INSTALL THE SOFTWARE .................................................................................................................................4
TROUBLESHOOTING ........................................................................................................................................4
GETTING STARTED .........................................................................................................................................5
4.1
MAIN PANEL OVERVIEW .................................................................................................................................5
4.2
PRACTICE WITH IMAGES AND MOVIES .............................................................................................................6
4.2.1
General Instructions ..............................................................................................................................6
4.2.2
BGA inspection ......................................................................................................................................6
4.2.3
Face recognition on soccer teams and the ferret database ...................................................................7
4.2.4
Road Marking detection.........................................................................................................................8
4.2.5
Inkjet character recognition ..................................................................................................................8
4.2.6
Glass surface anomaly detection ...........................................................................................................9
4.2.7
Wafer inspection ....................................................................................................................................9
4.2.8
Introduction to image compression with NeuroMem ........................................................................... 10
4.2.9
Kanji Character recognition ................................................................................................................ 10
4.3
START A NEW PROJECT ................................................................................................................................. 11
4.4
SPECIFICATION OVERVIEW ........................................................................................................................... 12
5
NAVIGATING THROUGH IMAGES ............................................................................................................ 15
6
NEW PROJECT SETTINGS ........................................................................................................................... 16
6.1
6.2
7
NAME YOUR CATEGORIES ............................................................................................................................. 16
REGION OF INTERESTED: SIZING AND FEATURE EXTRACTION ....................................................................... 17
TEACH EXAMPLES ........................................................................................................................................ 18
7.1
8
LEARN ANNOTATIONS .................................................................................................................................. 18
RECOGNITION ................................................................................................................................................ 18
8.1
9
COPING WITH CHANGE OF SCALE .................................................................................................................. 18
VIEW KNOWLEDGE CONTENT ................................................................................................................. 18
10
10.1
10.2
PROJECT MANAGEMENT........................................................................................................................ 19
COGNISIGHT PROJECT FILE........................................................................................................................... 19
PREFERENCE FILE ......................................................................................................................................... 19
General Vision
3
2
Introduction
2.1
What can I do with Image Knowledge Builder?
-
-
3
Easy navigation through folder of images and movie files
Build a knowledge to recognize objects
o Learning annotations edited by the user
o Learning labeled images
o Learning unsupervised
Generate outputs showing the recognized objects
o Visual color and text overlay
o Map of the hit locations
o Map of the category locations
o Map of confidence
o Text list of the recognized objects or anomalies
Installation
3.1
Preliminaries
If you intend to use the Image Knowledge Builder in conjunction with a device featuring a NeuroMem network,
start by installing the driver of your hardware if applicable.
3.2
Install the software
-
3.3
Double click at the setup.exe file and follow the instructions on the screen.
After the display of the panel “Installing GV Image Knowledge Builder…”, you may get a pop-up message
asking you to confirm the installation of the software from an Unknown Publisher. Click Yes to proceed.
Upon termination, click at Start menu \All Programs\General Vision\Image Knowledge Builder
Wait a few seconds for the program to launch
Troubleshooting
Error message: “Unable to load GVEngine.dll”
- Go to C://Program Files/General Vision/Image Knowledge Builder/ FTDI Windows Drivers
- Double click at dpinst-amd64.exe (or dpinst-x86.exe if you have a 32-bit machine) and follow the
instructions on the screen.
- Upon successful termination, click at Start menu \All Programs\General Vision\Image Knowledge Builder
- Wait a few seconds for the program to launch
General Vision
4
4
4.1
Getting Started
Main panel overview
General Vision
5
4.2
Practice with images and movies
Image Knowledge Builder is supplied with example projects and their accompanying images and videos. They are
usually organized per topic and the name of the project file matches the name of the folder where you will find the
images used for the training and the verification.
4.2.1
General Instructions
The proposed examples can be reviewed in a few minutes as follows:
-
4.2.2
Use the navigation bar to open an image stored in C:\\yourUser\\Documents\\General
Vision\\Images\\Topic.
Select File\Open Project from the Main menu and look for an existing *.prj file with the Topic label.
A summary of the project definition is displayed in the Status bar at the bottom of the screen: size of the
knowledge, type of feature extraction, image scanning mode.
If neurons are committed, their models can be viewed in the View/Knowledge menu, or you can click
directly at the Reco button to view the results of the recognition.
If no neurons are committed but annotations can be seen as overlay in the left image window, you can
click at the Learn button and when done at the Reco button
If no annotations exist for the selected image, you can edit your own manually, then click at the Learn
button and when done at the Reco button
BGA inspection
Example in machine vision where a non-linear classifier is a simple remedy to learn
a diversity of examples in a few mouse clicks. Try to achieve the same results in
less than a minute with a threshold technique!
1.
Click Browse in the Navigation toolbar and load the image stored in Images/BGA folder.
General Vision
6
a.
2.
3.
4.
4.2.3
Assuming that the Learn method is set to the default “Annotations”, the program should
automatically find and load the annotations associated to this image and stored in the file
BGA_GT.txt stored in the same folder.
b. Note that the annotations include “background” annotations (in black). They behave as counter
examples to ensure that the neurons learning solder balls (in red) do not over generalize too
much.
Select Load project in the File menu and load the file bga.prj stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is empty. The size of the
region to learn and recognize is 12x12 pixels and the feature extraction a monochrome and
subsample.
Checked the Auto checkbox to automatically recognize the image when a Learn operation or Load image
is executed. Verify that the Reco method is set to Objects.
Click Learn to learn the annotations seen on in the learning panel. Verify that all the solder balls are
highlighted in the right panel. Some appear more reddish than other when a solder ball is recognized at 2
adjacent positions.
Face recognition on soccer teams and the ferret database
A great example for cooperative face detection when people are facing forward.
Observe how 7 models can recognize soccer players from four different teams as
well as faces extracted from the ferret database!
1.
2.
3.
4.
5.
6.
7.
Click Browse in the Navigation toolbar and load the image stored in Images/Faces_SoccerPlayers folder
Select Load project in the File menu and load the file faces_soccerplayers_kn7.prj stored in the Projects
folder
a. In the status bar at the bottom of the screen, notice that the knowledge is already built and
consists of 7 neurons. The size of the region used to learn and recognize the characters is 27x 36
pixels. The feature extraction a monochrome normalized subsample.
The knowledge can be displayed by clicking “View Knowledge” button.
Click the Reco button to inspect the entire image. Note that this can take a few seconds in simulation.
Select the Auto checkbox next to the Reco button. Go to the next image by clicking at the “>” button in
the navigation bar. Review the results. Proceed to next image.
Using the navigation bar, select the first image supplied in the Images/faces_ferret folder. The Ferret
database is a known repository of faces looking at the camera from approximately the same distance. The
region of interest is clearly too small in this case to cover the eye and nose area of the person. In the
navigation bar, reduce the image by a scale 1/3 and you should observe that the face is then detected.
Go through several images of this same folder by clicking at the “>” button in the navigation bar.
General Vision
7
4.2.4
Road Marking detection
Monitor the edge and center lane while driving on Hwy 101
1.
2.
3.
4.
5.
4.2.5
Change the data source in the Navigation toolbar to work on “Movie” and load the RoadTrip.avi file stored
in Videos/RoadTrip folder
Select Load project in the File menu and load the roadtrip.prj file stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is empty. The size of the
region to learn and recognize as the lane marking is 24 x 8 pixels. The feature extraction a color
normalized subsample.
Click the Learn button to build the knowledge
Click the Reco button to inspect the entire image. Note that this can take a few seconds in simulation.
Select the Auto checkbox next to the Reco button. Go to the next image by clicking at the “>” button in
the navigation bar. Review the results. Proceed to next image.
Inkjet character recognition
This example quickly demonstrates how the feature extraction can impact the
ease of deployment of an application, in particular limit the number of examples
to teach without compromising the accuracy of the recognition.
1.
2.
3.
4.
5.
6.
Click Browse in the Navigation toolbar and load the first image stored in Images/OCR_inkjet folder:
Im_110217A_1008.jpg.
a. Assuming that the Learn method is set to the default “Annotations”, the program should
automatically find and load the annotations associated to this image and stored in the file
Im_110217A_1008_GT.txt stored in the same folder.
Select Load project in the File menu and load the file ocr_inkjet.prj stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is empty. The size of the
region to learn and recognize is 36x60 pixels and the feature extraction a monochrome and
normalized subsample.
Checked the Auto checkbox to automatically recognize the image when a Learn operation or Load image
is executed. Verify that the Reco method is set to Objects.
Click Learn to learn the annotations seen on in the learning panel. Verify that the digits are highlighted
with the proper categories in the right panel. The color legend can be seen and changed in the
Options/Settings menu.
Click at the Next arrow in the navigation bar to proceed with the next image. Notice that on the third
image, a new annotation appears for the digit “6” which has still not been taught. Click Learn to add it to
the current knowledge composed of 7 neurons as reported in the status bar. You can review the results
on the next and previous images navigating through the various images supplied in the OCR_inkjet folder.
Now, let’s repeat the same routine, but with a slightly different feature extraction:
General Vision
8
a.
b.
7.
4.2.6
Erase the knowledge by selecting the Clear command in the Knowledge menu
Open the Options/Settings panel and uncheck the “Normalize” checkbox. This means that the
amplitude of the subsample will not be expanded to the range [0, 255].
Load the first image again with its annotations, learn and review the results of the recognition:
a. Select Options/Details at cursor and move the cursor over the image
b. Observe the number of misfiring such as in (320,90)
c. Observe how the number of uncertainties such as in (54,70)
Glass surface anomaly detection
This example is a good illustration of the power of the NeuroMem RBF classifier
to detect novelty which in this case represents the occurrence of an anomaly.
The learning is unsupervised and consists of using the neurons to build a
codebook of good texture pattern from a reference image. In new images, the
program only reports the locations where no neuron fires, thus indicating an
anomaly in the glass pattern.
1.
2.
3.
4.
5.
6.
7.
4.2.7
Change the data source in the Navigation toolbar to “Image” and load the Glass0_good.png file stored in
Images/Surface_Glass folder
Select Load project in the File menu and load the surface_glass.prj file stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is empty. The size of the
glass area to learn and recognize is 16x16 pixels. The feature extraction a monochrome
subsample.
Select the “Codebook” learning mode to build the knowledge and click the Learn button. Using the Maxif
of 1500, and a learning step of 16 along the horizontal and vertical axes, 12 neurons are committed.
Select the “Anomalies” recognition mode, check the Auto checkbox and click the Reco button.
No anomaly is reported in the good image.
Go to the next image by clicking at the “>” button in the navigation bar. Anomalies should be reported in
the other three images.
You can practice with the same routine using bigger or smaller values of the maximum influence field and
observe the differences in the number of committed neurons and number of discrepancies identified in
recognition.
Wafer inspection
This example illustrates the use of an unsupervised learning method to learn each
block of a reference image such as a wafer, a printed material, etc and detect
significant local differences in other similar images.
1.
2.
Click Browse in the Navigation toolbar and load the image Good.tiff stored in Images/Wafers folder
Select Load project in the File menu and load the file wafer_inspection.prj stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is empty. The size of the
region to learn and recognize is 16x16 pixels and the feature extraction with the value 3 means
that the subsample includes the color information. The maximum influence field of 2000 is an
important parameter in this example and tunes the conservatism of the neurons.
General Vision
9
3.
4.
5.
6.
7.
4.2.8
Select the “Golden template” learning mode to build the knowledge and click the Learn button. Using the
Maxif of 2000, and a learning step of 16 along the horizontal and vertical axes, 674 neurons are
committed.
Select the “Anomalies” recognition mode, check the Auto checkbox and click the Reco button.
No anomaly is reported in the good image.
Go to the next image by clicking at the “>” button in the navigation bar. Anomalies should be reported in
the other three images.
You can practice with the same routine using bigger or smaller values of the maximum influence field and
observe the differences in the number of committed neurons and number of discrepancies identified in
recognition.
Introduction to image compression with NeuroMem
This example illustrates the use of an unsupervised learning method to build a
codebook of reference codes over one or multiple images, and reconstruct
simpler images by replacing the original blocks by their closest matches.
1.
2.
Click Browse in the Navigation toolbar and load the first image stored in Images/MiscBWImages folder
Select Load project in the File menu and load the file compression.prj stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is empty. The size of the
region to learn and recognize is 9x9 pixels and the feature extraction is a monochrome
subsample includes the color information.
3. Select the “Codebook” learning mode to build the knowledge and click the Learn button. Using the Maxif
of 2000, approximately 50 neurons are committed.
4. Select the “MapOfModels” recognition mode, check the Auto checkbox and click the Reco button.
5. A simplified image based of only 50 patches is generated.
6. Go to the next image by clicking at the “>” button in the navigation bar and observe the same
compression effect.
7. If blocks appear in black in an image, click Learn. A few neurons should be added to the knowledge. They
correspond to the blocks not recognized in the previous codebook.
8. You can practice with the same routine using smaller values of the maximum influence field and observe
the differences in the number of committed neurons and the reduction of the error between the source
and its Transform image.
9. You can practice with the same routine using smaller patches of pixels and observe a finer granularity and
lesser error in the Transform images.
10. You can practice with the same routine using the images of the MiscColorImages folder. Change the
feature to Color Subsample in the Options/Settings menu if you wish to build a codebook of color patches.
4.2.9
Kanji Character recognition
This example demonstrates that the NeuroMem network is (1) a good classifier
which can easily discriminate between many models, and also (2) the recognition
time is not impacted by the number of models which constitute the knowledge
base (360 in this case).
Warning! If using a simulation of the NeuroMem network, the recognition may
take several seconds. On the small image, the window of interest is moved across
General Vision
10
68,432 positions and at each position, the block of pixels (256 bytes) is compared
with 360 models. Assuming that each comparison between an input pixel and the
same pixel stored in a neuron takes 4 operations, the recognition of the entire
image is equivalent to 25.2 billions of operations. The CM1K chip will execute this
in 2 seconds or 12.5 Giga ops.
1.
2.
3.
4.
5.
6.
4.3
Click Browse in the Navigation toolbar and load the image stored in Images/Kanji folder
Select Load project in the File menu and load the file kanji360.prj stored in the Projects folder
a. In the status bar at the bottom of the screen, notice that the knowledge is already built and
consists of 360 neurons. The size of the region used to learn and recognize the characters is
16x16 pixels. The feature extraction with the value 0 corresponds to a monochrome subsample.
Select Options/Overview at Cursor. Move the cursor over a character and view the result of the
recognition in the Cursor Info box of the main panel. Note that positive recognition of a character only
occurs when well centered over it.
The knowledge can be displayed by clicking “View Knowledge” button. Note that in this case, each
character is represented by 3 models shifted by one pixel each. This training was done intentionally to
allow the detection of a character at different offsets.
Click the Reco button to inspect the entire image. Note that this can take a few seconds in simulation,
since the image is entirely scanned with a step of 1 pixel per column and 2 pixels per row, and at each
position the block of 16x16 pixels is matched against 360 models.
Review the results in the Options/Results window
Start a new project
A new project can be defined in 6 easy steps:
- Configure the camera (gain and shutter)
- Define the size of the objects to recognize (a part, a component within a part, a window, or else)
- Define the categories of objects to recognize (i.e. good, bad; or grade1, grade2, grade3; or A,B,C,D,E; etc)
- Teach examples for each category of objects
- View the knowledge generated by the neurons
- Monitor results on screen
- Collect images and save them to disk for further training and validation
- Save images based on the result of the recognition (i.e. selective recording)
For more information and application notes, please refer to the Resources page on our web site at www.generalvision.com.
General Vision
11
4.4
Specification overview
Version 2.0
Access
Next
revision
Image sources and navigation
Load image or movie file
Main/Navigation toolbar
Navigate through folder of the last selected image
Main/Navigation toolbar
Navigate through frames of the last selected video
Main/Navigation toolbar
Access to 1st, previous, next, last
Main/Navigation toolbar
User defined step
Options/Settings/Image
Handling
Main/Navigation toolbar
Play/Pause
Pre-processing
Image scaling
Main/Navigation toolbar
Report of the scaling factor and resulting image resolution
Main/Status bar
Learning functions and utilities
User defined feature extraction and global network settings
Choice of feature extraction on color or monochrome components
of an image
Choice of subsample or histogram feature vector, normalized On or
Off
Selection of the learning method
Options/Settings/Feature
Extraction
Options/Settings/Feature
Extraction
Options/Settings/Feature
Extraction
Main/Learning combo box
Learn annotations shown on screen
Main/Learn button
Learn an image obtaining input categories from an 8-bit GroundTruth image
Build codebook from image
Main/Learn button
Learn golden template (codebook and list of codeIDs)
Main/Learn button
Report committed neurons
Main/Status bar
Report selected feature extraction and nbr of annotations, if
applicable
Undo last learning
Main/Status bar
Load Ground truth image
Main/File/Load GT image
Main/Learn button
Knowledge/Undo menu
Learn an ROS
x
Build codebook over an ROS
x
Batch learning of annotated images
x
Learning mode for exhaustive annotations (every thing else
becomes background)
x
Image annotation
User defined names and color tags to annotate categories of objects
Options/Settings/Categories
Annotate a single region
Mouse click with right button
Paint to annotate a surface
Mouse move
button down
General Vision
holding
right
12
Delete annotation
Report nbr of annotations, if applicable
Save annotations of an image (default name is image name +
"_GT.txt")
Save annotations of a video frame (default name is moviename+
_frame number + "_GT.txt")
Automatically load an annotation file, if it exists, when opening a
new image
Automatic addition of annotations as counter examples
Shift key + Mouse move holding rioght
button down
Main/Status bar
Main/File menu
Main/File menu
Main/File menu
x
Automatic addition of annotations as adjacent examples
x
Local recognition functions and utilities
Display feature vector at cursor position
Options/Overview at Cursor
Report ground truth and recognized category, if applicable
Options/Overview at Cursor
Report feature vector, firing models and ground truth, if applicable
Options/Details at Cursor
Image recognition functions and utilities
Selection of the recognition method
Main/Recognition combo box
User defined scanning options (step and skip)
Options/Settings/Scanning
options
Options/Settings/Categories
User defined names and color tags to annotate categories of objects
Find all recognized objects (visual tag)
Find anomalies or non recognized objects (visual tag)
Find differences (requires that the learning method is Lean
Template)
Report X,Y locations and categories in a gridview
Main/Reco button
checkbox
Main/Reco button
checkbox
Main/Reco button
checkbox
Options/Results
Report the sequence of recognized categories in a string (useful for
OCR for example)
Select an X,Y location in the report to locate in the image window
Options/Results
Save to file X,Y locations and categories
Options/Results
Generate map of categories (from CAT, useful for 2nd level reco)
Main/Reco button
checkbox
Main/Reco button
checkbox
Main/Reco button
checkbox
Main/Auto checkbox
Generate map of confidence (from DIST, useful for 2nd level reco)
Generate map of identifiers (from codeIDs, useful for template
matching)
Auto checkbox to recognize each time a new image/frame is loaded
Report selected scanning options and nbr of detected objects if
applicable
Generated maps can be saved and become Ground Truth template
images
Report of the color palette used to display the map on screen
or
Auto
or
Auto
or
Auto
Options/Results/Click at a cell
or
Auto
or
Auto
or
Auto
Main/Status bar
Main/File/Save Transform menu
Main
Recognition limited to 1 or multiple ROS
x
Locate specific categories
x
Generate a log of images containing specific categories
x
General Vision
13
Knowledge display and file management
Display all models in a tumbnail
Knowledge/View
Display models per category in a thumbnail
Knowledge/View
Display all neurons' content in a gridview
Knowledge/View/Details tab
Clear all neurons
Knowledge/Clear
Restore the knowledge as it was before the last learning command
Knowledge/Undo last learning
Save to /Restore from file
Main/File menu
Choice of the neuron platform (simulation or hardware)
Knowledge/Platform menu
Outputs
Save the knowledge and settings used to build it
Main/File menu
Transform image such as the map of categories and confidence
Main/File menu
Annotation file associated to an image file or frame in a movie file
Main/File menu
Text file with X,Y, Category of recognized objects in an image
Options/Results
Text file with X,Y, Category of recognized objects in a movie file
x
Other
Options to save the last project at default for the next launch
Main/File menu
Ready to use examples with images and pre-defined projects and
annotation files
Example BGA
Main/File menu
Example Inkjet OCR
Main/File menu
Example Kanji
Main/File menu
Example Face Reco on soccer teams
Main/File menu
Example glass surface anomaly detection
Main/File menu
Example wafer inspection w. template matching
Main/File menu
General Vision
Main/File menu
14
5
Navigating through images
-
Load images from a folder
Load movie files
Go to 1st, previous, next or last image in folder
Go to 1st, previous, next or last frame in movie
Go to specific image in a folder (edit value and type Enter, or use the scrollbar)
Go to a specific frame in a movie (edit value and type Enter, or use the scrollbar)
Scale the image by zooming out
General Vision
15
6
6.1
New project settings
Name your categories
The Teach button opens the following panel. The first step consists of defining the categories of objects you want
to discriminate.
The names of the Category buttons can be changed by clicking at the corresponding button with the right mouse
button. An Edit box is prompted and a new name can be edited. Make sure to press enter to validate the new
name.
Examples of categories for a part inspection
- Acceptable (category 1)
- Recyclable (category 2)
Or
- Grade A (category 1)
- Grade B (category 2)
- Grade C (category 3)
Examples of categories for a traffic monitoring application:
- Car (category 1)
- Truck (category 2)
- Van (category 3)
General Vision
16
The color associated to a category can also be changed by clicking at the current color with the right mouse button.
A color dialog appears so you can assign a new color to the category button.
6.2
Region of interested: Sizing and Feature extraction
The Region Of Interest (ROI) is the smallest rectangular area including your object. It should be limited to the
smallest discriminant portion of your object and exclude any unnecessary background area.
For example, if a mechanical part can be classified as good or defective based on the presence or absence of a
screw at a given position in the part, the region of interest should be not be defined as the whole part but rather
limited to the rectangle where the screw is expected.
The feature extraction is the method used to convert the pixel data within the region of interest into a vector of
256 bytes. Standard methods include a subsampling and histogram calculation. In both cases, these signatures can
be based on the intensity value or the RGB values of the pixels. Finally, you can decide, or not, to normalize the
amplitude of the feature vector.
Remark: If you select a feature extraction based on a subsampling, the option “Details at cursor” (under the
Options menu) will display the feature as a 2D image.
General Vision
17
7
Teach examples
The IKB offers severalmmethods to train the neurons:
- Learn annotations
- Build codebook
7.1
Learn Annotations
Select the “Annotations” learning mode
Select the category of the object to annotate
Move the cursor over the object and click with the right mouse button
You can choose to learn at any time by clicking the Learn button, but it is recommended to first annotate all the
regions of interest and learn them all at once.
The annotation of an xyz image are automatically saved to a file with the extension xyz_GT.txt when you click the
Learn button.
If the “Auto” checkbox next to the Learn button is marked, the annotations loaded with an image are automatically
learned. Combine this with the Play mode of the navigation bar and you can learn automatically all the “annotated
images within a given folder.
8
8.1
Recognition
Coping with change of scale
Alternative #1:
- Learn at a given ROI size (as specified in the Options/Settings menu)
- In recognition, the image scale can be changed in the navigation bar at the bottom of the screen.
The project faces_soccerplayer is a good illustration of this approach: the ROI has been sized to learn on the
images of the soccer players. However, if you load an image from the folder faces_Ferret, you can notice that the
ROI size is too small, but if you scale down the image by 3, then it becomes suitable and the face is recognized.
Refer to manual paragraph 4.2.3.
Alternative #2:
- If the feature extraction can be a subsampling, it is scale invariant if the number of blocks within the ROI
remain the same. It is then possible to learn objects at different compatible ROI sizes (as specified in the
Options/Settings menu). For example an ROI of 64x64 with blocks of 4x4 creates a vector of 16x16
components, but similarly an ROI of 128x128 with blocks of 8x8 also creates a vector of 16x16
components.
- In recognition, the neurons holding models of the objects at the proper scale will fire.
9
View Knowledge Content
General Vision
18
The View Knowledge menu allows reviewing the contents of the neurons, the pattern they hold, their category
(Cat), and active influence field (AIF). This utility can be helpful to understand why some erroneous or uncertain
classifications occur.
The AIF is a dynamic attribute entirely controlled by the neuron itself. From the time the neuron is created, its
influence field can only reduce as new examples are taught and the knowledge gets more accurate. The smaller
the influence field the more conservative the neuron.
10 Project Management
Projects defined under Image_Knowledge_Builder can be saved to disk in a project file or to the Flash memory of
the camera.
A project is composed of two files associated to the same rootname:
- A project file (*.csp) containing the description of the recognition engine
- A settings file (*.prf) with the description of the GUI preferences
10.1 CogniSight Project file
-
The *.csp file describes the recognition engine. It contains all the information necessary to resume the
recognition of the images on an embedded system including the knowledge, the region of interest, the
region of search and sensor settings
-
10.2 Preference file
-
*_settings.csv describes the settings of the graphical user interface including the names and colors
associated to the category values, as well as the preferred learning and recording settings.
General Vision
19
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising