Deep learning segmentation

Object detection using pre-trained algorithms via ONNX models. The object detection algorithms supported can be found under the parameter model type.

Note that instance segmentation where individual pixels are masked is not supported.

See ONNX image segmentation for more information.

Parameters

Model type

The .onnx model type. Available options are:

• Faster R-CNN

• YOLO v4

• YOLO v5

• YOLO v8

• YOLO v11

Onnx file

Select the pre-trained ONNX file for the selected model type.

Source

On which image the ONNX segmentation should be applied. The pseudo-rgb image or a painted prediction image.

Confidence

The confidence level required by the model for an object to be categorized.

Normalize the pixel values

If Normalize and center is used the values will be scaled to 0-1 before the is subtracted and the used in the division. See below.

• No normalization

• Normalize and center

Image dimension order

In which order the input dimensions are:

• Width / Height

• Height / Width

Output layer to use

Which type of output layer to use:

• Sigmoid layer

• Detection layer

Intensity threshold

Filter grayscale values outside of interval.

Activation function for scores

Apply function to score values before applying the confidence value threshold

• None

• Sigmoid

• Tanh

• ReLU

Min area

The minimum number of pixels for an object to be included.

Max area

The maximum number of pixels for an object to be included.

Object filter

Use an expression to further exclude unwanted objects based on shape.

Operators than can be used expressions include the data operators wNNN and bMMM for referring to wavelength bands, the range operator : used for averaging data, standard arithmetic (+,-,/,* …) and comparison operators (=,>,< …) as well as some mathematical functions ( …) and constants ( ).

Properties that can be used for the Expression:

• Area

• Length

• Width

• Circumference

• Regularity

• Roundness

• Angle

• D1

• D2

• X

• Y

• MaxBorderDistance

• BoundingBoxArea

For details on each available property see: Object properties Details

Shrink

Takes away x numbers of pixels at the borders of the objects included in images.

Separate

The Separate parameter defines how pixels that pass your segmentation threshold are grouped into distinct object samples.

Pixel Connectivity Rule

Breeze uses 8-connectivity. Pixels are considered connected if they touch horizontally, vertically, or diagonally on a single corner.

Parameter Options

1. Normal

• Behavior: Standard connected-component grouping.

• Result: Pixels touching by an edge or corner form a single object. Physically isolated clusters form separate objects.

2. Separate adjacent objects

• Behavior: Applies a distance transform to find the inner centers (geometric peaks) of pixel clusters, then splits them using a watershed-style growth algorithm.

• Best Used For: Round or spherical objects that are touching or overlapping.

• Limitation: Elongated, branched, or irregular shapes may be incorrectly split into multiple objects because the algorithm detects multiple internal peaks within a single physical item.

3. Merge all objects into one

• Behavior: Disregards spacing; combines all valid pixels across the entire image.

• Result: Evaluates the entire frame as one single object for collective regional analysis.

4. Merge all objects per row

• Behavior: Groups detected objects based on the spatial grid of their vertical center points ($Y$-coordinates).

• Result: All objects residing within the same horizontal row lane are merged into a single object.

5. Merge all objects per column

• Behavior: Groups detected objects based on the spatial grid of their horizontal center points ($X$-coordinates).

• Result: All objects residing within the same vertical column lane are merged into a single object.

Max objects

Max number of objects in image, takes the first objects sorted by confidence.

Inverse

✅ Includes the opposite of the sample specified in the Deep learning image model.

⬜ Includes the sample specified from the Deep learning image model.

Link

Link output objects from two or more segmentations to top segmentation. Descriptors can then be added to the common object output and will be calculated for objects from all segmentations.