2D Preprocessing Function Configuration Guide
About 4555 wordsAbout 15 min
2D preprocessing takes place before Instance Segmentation. It improves input images, brings out target features, or reduces the effect of irrelevant regions on recognition results. This page first helps you choose functions based on the issue at your site, then explains their parameters and tuning methods in the order they appear in PickWiz.
Where to configure
In Visual Parameters, go to 2D Recognition > Preprocessing, add the required functions, and set their parameters. Run with the default values first, then adjust one parameter at a time based on the results.
Choose a 2D preprocessing function
Identify the main issue in the original image, then choose the corresponding function. Avoid stacking functions simply to add processing steps.
| Issue at the site or processing goal | Recommended function | When to use it |
|---|---|---|
| Noise reduction and smoothing: the image has local depth fluctuations | Bilateral Filter | Smooths the image. While tuning, check that target edges and small structures remain clear. |
| Depth image processing: a per-pixel normal vector image is needed | Depth to Normal Vector Image | Calculates per-pixel normal vectors from the depth image and outputs a normal vector image. |
| Brightness and contrast: saturation, contrast, brightness, or sharpness needs adjustment | Image Enhancement | Lets you select an image property and set the degree of enhancement. |
| Brightness and contrast: foreground and background lack contrast | Histogram Equalization | Improves image contrast through local or global histogram equalization. |
| Depth image processing: depth regions must be retained or excluded based on color | Filter Depth Image by Color | Uses an HSV color range to select regions and determines whether to retain depth data inside or outside that range. |
| Brightness and contrast: the entire image is too bright or too dark | Gamma Image Correction | Adjusts image brightness using Gamma coefficients. |
| Depth image processing: target regions contain depth holes | Fill Depth Image Holes | Fills missing regions in the depth image and smooths the result. |
| Region and top layer extraction: targets overlap or occlude one another, making their edges unclear | Edge Enhancement | Highlights target edges based on normal vector direction. Suitable for flat targets and curved targets such as sacks. |
| Region and top layer extraction: the texture of the topmost or bottommost targets must be retained | Extract Top Layer Texture | Commonly used when depalletizing cartons one at a time. Reduces false detections caused by similar textures on upper and lower layers. |
| Region and top layer extraction: background outside ROI3D affects detection | Remove Image Background Outside ROI3D | Sets regions outside ROI3D to a background color or a contrasting background color. |
| Depth image processing: a subsequent Point Cloud process needs normal vectors | Calculate Normal Vectors | Calculates Point Cloud normal vectors for subsequent Point Cloud processing. |
| Region and top layer extraction: redundant background Point Cloud data must be filtered out | Extract Top Layer Point Cloud | Retains the top layer Point Cloud based on the depth image, reducing interference from background Point Cloud data. |
Tuning principles
- Check camera exposure, lighting, and image acquisition quality first.
- Change only one parameter at a time, and record the results before and after each change.
- Evaluate preprocessing results together with subsequent Instance Segmentation results; do not judge them from a single image alone.
- Verify the settings with samples from different batches and positions so they do not work for only one view.
2D preprocessing functions
The functions below follow the order of the function list in PickWiz so you can compare them with the software configuration.
Related documents
Continue configuring the recognition and segmentation process
Select Instances for subsequent processing
Set the priority order for processing Instances
Generate the Point Cloud required for subsequent 3D calculations