| |
 |
Learning
a Model |
- Target: a target
is a form that you wish to recognize (e.g. a car, a face,
a weapon…). SpikeNet allows you to search for multiple
targets simultaneously.
- Input image:
The Input Image is the image on which your perform the search
using the models in the database.
- Source image: Image used to create a
model. You can use the entire image or only a selection.
- Region: You can
use various tools to select a region for learning. The default
selection tool is rectangle but other tools can be selected
using the Region Selection Tool option of the Edit menu:
rounded rectangle, ellipse, free-hand. You can also fix
a determined size using the select fixed size option in
the Edit menu.
- Learning a model: Process to create models.
Select a region then double click within that region. The
model is automatically added to the model list and displayed
in the “Model” window.
- Model: a model is a region (an object,
a face, a car, a wing of a plane…) selected from a
source image that you wish to locate on other images. You
can create any number of models for each target that you
wish to recognize and add them to the Model List. You can
save these models in a Models File (.snm) which can be loaded
subsequently and used in the recognition.
- Models file: File containing models (.snm)
- Model Identificator (Model ID): Identificator
used for specifying a model within the database. It is unique
and given by the system during the addition of the model
to the database.
- Model Logical Name:
Name given to a model. By default, the name is that of the
source image file used to create the model. If you create
a model using a source image file named “airplane”,
the first model you create will be named “airplane
1”, the second, “airplane 2”, and so on.
Each Model has a unique Logical Name which you can change
by editing the Model Parameters.
- Model Generator:
When you create a model, you can use the Model Generator
to specify a range of orientations and zoom factors, together
with the step size between models. Click on the Learn option
in the Model menu. By default the Zoom and Rotate factors
boxes are left unchecked. If you check them, SNVision MB
will generate multiple models to cover the specified range
of zoom factors and/or orientations.
- Model Info: Properties of a model. Model
properties can be accessed by clicking on the "+"
sign to the left of the model name in the Model List Window.They
include:
- Model Identificator
- Model size: width and height of the model
- Heigh and width references (size of the original selection
on the source image)
- Model Logical Name
- Color: A colored shape (circle, cross, point, zone) is
used to locate the model when it is recognized on an image.
This color can be changed when you edit the model.
- Zoom: zoom factor applied
- Rotate: rotation factor applied
- Threshold: threshold value
- Detail: detail value
- Model List: Window displaying all models
by name. The checked models are enabled.
- Model parameters: Default search settings
are assigned to the model that you learn and then search.
Model search settings can affect the speed and robustness
of your application. It is recommended that you begin with
the default settings, and then adjust the individual settings
one by one in order to train the system to search for the
model in the most efficient manner.
- Threshold: Adjusts
the sensitivity of the model. Increasing the threshold will
increase the sensitivity of the model, generating hits that
better match your model. The higher the threshold value
is, the more similar to the model an object has to be in
order to be detected
Lowering the threshold will lower the sensitivity increasing
the amount of error. This can be useful when you have used
a single model (a face for example) and want the system
to respond even when the view is not identical to that of
the model used.
- Detail: Quantity
of information contained in a model. Increasing the detail
value will improve the signal to noise ratio, but with the
penalty of increasing the processing time.
- Enabled (Model): By default all models
are enabled. But you can disable models within your list.
This can be useful if you have created models for different
targets and want the recognition process to focus only on
specific targets.
- Color: Color of the shape used to locate
the model on the input image.
- Name: Model name
|
Recognizing a model |
- Recognition: Searches for model(s) in
the Input image and writes the results in a Results table.
The recognition process will find all the occurrences of
each enabled model. But, you can limit the number of hits
(detections) produced. Click on the “Recognize”
button to start the search.
- Recognition Parameters
Default settings are assigned when you first initiate the
recognition process. You can adjust all recognition parameters
in order to improve the performance and robustness of the
recognition.
- Spatial filter: Filter used to prevent
multiple recognitions within the same region. Only the detection
with the higher Quality score remains.
- Propagation: information extracted from
the input image.
When you increase the propagation value, more information
about the image is provided thus improving the signal to
noise ratio of the recognition process. But as more information
is provided, the processing time goes up and you increase
the number of false positives. The propagation value should
be kept below 30.
- Minimum Contrast:
SpikeNet’s algorithm is designed to be extremely good
at detecting objects even when the contrast in the image
is very low. One problem that this can cause is that in
an image where the only interesting objects are at high
contrasts, unwanted detections may occur in regions of very
low contrast. To avoid this sort of problem, you can raise
the Minimum Contrast value that is normally fixed at 0.
This rules out the false detections occurrences in regions
with very low contrast and greatly speeds up the processing
time. Only the regions in which the contrast is above the
Minimum Contrast value are subject to processing.
- Resolution:
The resolution refers to the spacing of the pixels in an
image. The higher the resolution, the more pixels in the
image. Decreasing the resolution will improve the processing
time and strengthen the robustness of the recognition because
you reduce the image detail. Make sure that you modify the
image resolution before learning a model.
- Maximum hits: When you start a recognition
process, you can set the expected number of occurrences
of the model to be found in the Input Image and shown in
the Recognition Results Table. The Maximum Hits value applies
to all models in the list.
- Multiscale: The multiscale option enables
you to process an image at several different scales simultaneously.
This provides an alternative way of achieving size invariant
recognition to that provided by the zoom options that can
be used during model learning (see Model Generator).
- Mask: Image used for
defining a region of interest within an image thus reducing
the processing time and the number of false positives in
the hidden regions (outside the mask).
To create a mask, select a region using the Region Selection
Tools in the Edit menu, and click on Add mask in the Processing
menu. You may save the image mask for later use (.pgm file)
by clicking on Save Mask in the Processing Menu.
|
Recognition Results |
- Hits: Occurrences (detections) of the
enabled model(s) found on an input image
- Results: Each time the recognition process
is initiated the Results window provides the following information
about the recognition:
- Rank: Detections (hits) listed by descending
order of quality
- ID: Model Identificator
- Model Logical Name
- Model X and Y coordinates: When an occurrence
of the model is found, the X and Y coordinates of the actual
found position are provided. The reference point is at the
center of the model.
- Quality: The quality
is a confidence factor. Quality scores range from 0 to 2000
with higher values corresponding to higher quality.
- Frame: Image number in a video sequence.
The status bar of the main window also provides information
on the recognition results:
- Hit Count: Number of hits that
were found in the active Input Document (Image, Video
Sequence, Video Capture).
- Processing time in ms
|
| |
|
|
|
