Sheldon L. Epstein
Recent advances in digital image cameras have provided new opportunities for engineering machine-vision systems for automatic parts inspection as well as new demands on imaging software. Digital cameras can generate images containing several millions of pixels, and they can transmit them at rates as high as 100 Mpixel/s, which puts a premium on efficient imaging-algorithm design.
Some of the new digital cameras also have finer pixel pitches and higher pixel resolution than earlier models. Engineers now want better morphology algorithms and want to work with 12- to 16-bit data words.
One example of a new opportunity made possible with digital cameras is the examination of concave surfaces such as those of an automotive shaft seal (see Fig. 1). Using a custom optics and illumination system, including a digital camera (Dalsa; Waterloo, Ontario, Canada), a PIXCI Model D imaging board designed for that camera, and intelligent XCAP imaging software (EPIX; Buffalo Grove, IL) that detects the camera on startup and configures itself appropriately, we have been able to capture images and format them for automatic inspection (see Fig. 2).1 Manufacturing defects, such as a bent seal frame, are readily detected and measured using the system, and other defects, such as molding flash, cuts, missing grease, and eccentricity, are identified.
Shaft seals are excellent examples of products that can be more readily examined by machine than by hand because their performance depends on the quality of their concave surfaces. The new higher-resolution digital cameras make this practical. The process of inspecting them consists of eight steps. Almost all of the options described in steps 2 to 8 are available in imaging software for digital cameras.
Inspection process
First, several images of the concave surface are acquired. Each is taken at a different rotational orientation so that the entire 360° interior surface is captured. Next, the individual images are mapped from camera space to analysis space. Depending on the camera and lens used, the individual images may need to be warped or mapped so that curvilinear lines, such as those shown in Fig. 1, become straight lines. This is because it is much easier and faster to analyze straight lines.
The third step is to tile the individual images that have been mapped into analysis space to construct a single image of the entire concave surface. One method of accomplishing this is to open a null-destination image frame of a size at least equal to the entire concave surface and then to copy each of the individual images into the destination frame. Copy and resize functions permit precise alignment and adjustment to cancel out small differences in the individual images. The result is a single tiled image of the entire 360° interior surface.
Next the destination image frame must be normalized for the purpose of canceling slight differences in density between the individual images. Normalization functions examine the statistics of each row or column of the destination image frame and rewrite each of the rows or columns so that the destination image appears to have been originally captured as a single frame. This step is necessary to avoid the mistaken identification of imaging artifacts as product features.
Contrast enhancement follows. Most images of machined parts have a narrow distribution of pixel densities in areas of special interest. This is sometimes caused by the need to keep spectral reflections from saturating the camera sensor. To provide a wider distribution of pixel values, contrast-enhancement algorithms can be used to remap pixel densities in the entire destination image frame or in one or more areas of interest. There are several options, including histogram equalization, contrast inversion, and contrast stretching. In some situations, contrast enhancement should occur before normalization.
Feature detection and analysis
The sixth step is feature detection. For products such as shaft seals, the principal features of interest are edges. There are three families of edge-detection algorithms. The first type of edge detector depends on the magnitude of the density of pixels on edges. These algorithms are the fastest and work well when the edges are straight and parallel with the destination frame border. A second type of edge detector depends on the gradient of pixel densities in directions parallel to the destination frame border sides. They execute slower than magnitude detectors and also work well when the edges are straight and parallel with the destination frame border sides. Their advantage is that they can detect edges of rounded surfaces more reliably than magnitude detectors.
The third type of edge detector is based on morphology (calculus of shapes) mathematics. One concept is to select a small kernel that can be used to erode the image by making features smaller. If the eroded image is then subtracted from its original image, the resulting difference image will contain edges along the locus of the erosion. Morphological edge detectors are slow in execution; however, they are preferred when the image contains curvilinear edges. Another type of edge detector is the family of specialized transform functions that are used to identify lines, circles, and ellipses. One popular function is the Hough transform that can return the slopes and offsets of straight lines in an image irrespective of orientation with respect to frame boundaries.
The seventh step is feature analysis. Once features, such as edges, have been detected, they must be measured and analyzed. For example, a system can test to determine whether an edge is straight. In the case of the shaft-seal image, which contains several horizontal edges running parallel to the entire horizontal sides of the destination frame, the lines are not straight. The presence of curvilinear segments means that the seal's frame is bent.
The last step is statistical analysis. Detecting faults, such as bent frames, should not be the endpoint of image analysis. Because the loci and sizes of defective features are now known, databases can be constructed for statistical analyses. These analyses can discriminate between random events and systematic problems in a production process.
New digital cameras expand opportunities for machine vision and image analysis by providing higher resolution, more flexibility, and faster acquisition. One problem is that these cameras have complex interface requirementsoften including a dozen or more parameter selections.
Another difficulty is that they can rapidly generate huge amounts of image data. For these reasons, hardware and software selections are critical decisions. Custom interface boards and intelligent software save developer time by being able to run these cameras right out of the box. Advanced software can analyze large image data files for detection of important features.
REFERENCE
See the following Web sites:
- www.epixinc.com/pixci_d.htm;
- www.epixinc.com/xcap.htm; and
- www.k9ape.com/new/MSeal.html
SHELDON L. EPSTEIN is the chief engineer at Epstein AssociatesK9APE, POB 400; Wilmette, IL 60091-0400; e-mail: [email protected].