Innovative Production Machines and Systems Conference - neural networks

Intelligent Systems Tutorials

Vlado — Fri, 15 Jul 2005 16:14:14 +0100

Neural Networks in Automated Visual Inspection of Manufacturing Parts

Liatsis — Tue, 05 Jul 2005 10:38:49 +0100

Title slide only with powerpoint attachment

Some examples

gonzalo — Mon, 27 Jun 2005 18:17:29 +0100

Boxes determined by FMMIS on image samples for each of the 10 defect categories, from left to right and top to bottom: birdseye; blue stain; pocket; pith; wane; dead knot; split; live knot; stain; and hole.

Conclusions

gonzalo — Fri, 24 Jun 2005 16:52:24 +0100

&nbsp

Experimental results

gonzalo — Fri, 24 Jun 2005 16:50:50 +0100

The FMMIS global performance on the test set, and its per-category disaggregation are shown in the Table. The global TP rate achieved 95 percent of the total defects present in the test set, while the FP rate was 6 percent of the total grain lines (clear wood) present in the test set.
On average, the number of selected seeds per image was about 100, i.e., 0.1 percent of the total number of pixels of an image. This fact makes the FMMIS algorithm very fast.
The last column shows the average processing time for FMMIS, including the seed selection process, which reached 0.11 +/- 0.04 seconds per image.

Methods

gonzalo — Fri, 24 Jun 2005 16:48:03 +0100

A sample of 900 colour images (320x 240 pixels) of wood boards was drawn from the University of Chile. Each image was manually labelled according to its largest defect, into one of the following 10 defect categories (following the illustration from left to right, top to bottom): birdseye & freckle, bark & pitch pockets, wane, split, stain, blue stain, pith, dead knot, live knot and hole. The data set, was partitioned into two sets: 600 images for the training set and 300 images for the test set. The performance of the FMMIS algorithm was measured on the test set using the following criteria: number of true positives TP (the number of defects contained by hyperboxes and correctly detected), number of false positives FP (the number of grain lines contained by hyperboxes, i.e., detected as defects), and the average processing time. The area recognition rate (ARR) criterion allows to compare the area of the hyperbox built automatically by FMMIS with the area of the manually placed minimum bounding rectangle.

Output of the FMMIS

gonzalo — Fri, 24 Jun 2005 16:47:47 +0100

The last stage is to draw the rectangle (minimum bounding rectangle) on each defect using the min- and max-points of each hyperbox formed by the FMMIS algorithm.

FMMIS learning algorithm

gonzalo — Fri, 24 Jun 2005 16:45:11 +0100

1. Initialisation: All the min-max points are initially set to 0. The first pattern (seed) committed to a hyperbox results in a single point that is identical to the input pattern. 2. Hyperbox Expansion: Identify the hyperbox closest to the input pattern that can be expanded and expand it. If an expandable hyperbox cannot be found, add a new hyperbox. A fuzzy colour homogeneity criterion is defined to compare the colour similarity of two hyperboxes, then a user-defined parameter is introduced to control the required degree of colour homogeneity for expanding hyperboxes. 3. Hyperbox Ovelap Test: Determine whether the recent expansion caused any overlap between hyperboxes. 4. Hyperbox Contraction: If the expansion test identified any overlapping hyperboxes, contract the hyperboxes to eliminate the overlap. 5. Fine-tuning Hyperbox Expansion: Allows the hyperbox to grow if necessary until the defect is completely enclosed. 6. Hyperbox Merging: Process that merges hyperboxes belonging to the same defect, to ensure that each defect present in the image is contained by only one hyperbox.

Fuzzy min-max neural network for image segmentation (FMMIS)

gonzalo — Fri, 24 Jun 2005 16:39:42 +0100

&nbsp

Input patterns

gonzalo — Fri, 24 Jun 2005 16:33:51 +0100

&nbsp