Visual Inspection Product reliability is of maximum importance in most mass-production facilities.  100% inspection of all parts, subassemblies, and finished products. Therefore, the inspection process is normally the largest single cost in manufacturing. Most difficult task for inspection is that of inspecting for visual appearance.  Visual inspection seeks to identify both functional and cosmetic defects. Visual inspection in most cases depends mainly on human inspectors. Slide 2 Automated Visual Inspection Using Inductive Learning

Automated Visual Inspection Human inspectors are slow compared to modern production rates, and they make many errors. Automated visual inspection (AVI) is obviously the alternative to the human inspector. Several practical reasons for automated inspection include:  Freeing humans from dull and routine.  Saving human labor costs.  Performing inspection in unfavorable environments.  Reducing demand for highly skilled human inspectors.  Analyzing statistics on test information and keeping records for management decisions.  Matching high-speed production with high-speed inspection. Slide 3 Automated Visual Inspection Using Inductive Learning

Visual Inspection Techniques There are many techniques for automated visual inspection:  Image subtraction: ─ The inspected image to be is scanned and compared against the original image, which has been stored before. ─ The subtracted image is analyzed. ─ This method needs large reference data storage, accurate alignment, sensitive lighting and scanner conditions. ─ Also many images may not match point-by-point identically even when they are acceptable.  Dimensional verification: ─ The distance between edges of geometric shapes is the primary feature of this inspection method. ─ The task is to make a determination for each measurement as to weather it falls within the previously established standards. Slide 4 Automated Visual Inspection Using Inductive Learning

Visual Inspection Techniques (cont.)  Syntactic approach: ─ Uses descriptions of a large set of complex objects using small sets of simple pattern primitive and structural rules. ─ Primitives are small number of unique elements, as lines or corners. ─ A structural description of the primitives and the relationships between them can be determined to form a string grammar.  Feature (Template) Matching: ─ The inspected image is scanned and the required features are extracted. ─ Then these features are compared with those defined for the perfect pattern. ─ This method greatly compresses the image data for storage and reduces the sensitivity of the input intensity data. ─ A number of predefined binary templates can be used to extract the necessary features for images to be inspected. Slide 5 Automated Visual Inspection Using Inductive Learning

Template Matching Technique Mask technique can be used, with number of predefined binary templates, to extract the necessary features for inspected images. The total number of 3x3 mask templates is 28.  This number is calculated as follows: The total number of black pixels in each mask is 3.  The reason of using 8 is that, the central pixel is always black.  The rest of 8 pixels only 2 pixels can be black and the rest must be white. Slide 6 Automated Visual Inspection Using Inductive Learning Number of masks =

28 of 3x3 Masks Automated Visual Inspection Using Inductive Learning Slide 7

Template Matching Technique (cont.) The reason for choosing 3x3 masks is to reduce the processing time.  It is possible to have 5x5, 7x7 or some other masks.  If the size of the mask is bigger the accuracy may increase but the processing time will also increase. All 28 masks may not always be required to use for the applications.  The experience from many applications shows that, 10 to 15 masks are good enough to be employed. How to select the suitable masks for each application is an important problem. Slide 8 Automated Visual Inspection Using Inductive Learning

28 of 3x3 Masks Automated Visual Inspection Using Inductive Learning Slide 9 R R R R R R R R R R R

20 of 3x3 Masks Automated Visual Inspection Using Inductive Learning Slide 10

Mask Selection In order to select a proper number of masks, the following steps can be considered :  Select a number of example images.  Apply 28 masks and calculate the frequency of each.  Find the average frequency of each mask.  Sort the masks according to their average frequencies (from biggest to smallest).  Choose a number of them for the application. Each mask must be applied to each image pixel-by-pixel from left to right and from top to bottom. The frequencies may change from one image to another. We can take the average of all frequencies for the same mask and consider it for selection. Automated Visual Inspection Using Inductive Learning Slide 11

Inductive Learning Induction can be considered as the process of generalizing a procedural description from presented or observed examples. Inductive inference is the method of moving from specific examples to general rules. One of the visual pattern recognition goals is developing a system that can learn to classify patterns.  First, the system should be trained using a set of training examples.  Then it should use knowledge gained in the training session to automatically classify new examples. Automated Visual Inspection Using Inductive Learning Slide 12

Training Session A training process proceeds follows:  A number of good parts (examples) are shown to the system.  The frequencies of 20 3x3- masks are calculated.  Then an induction algorithm is used to extract the necessary rules. ─ The extracted set of rules represents the good parts.  When a pattern is shown to the system, using the extracted rules, it can decide whether it is good.  If the pattern cannot be decided as good it means that the pattern is bad (defected). ─ The system does not need to learn bad patterns. Automated Visual Inspection Using Inductive Learning Slide 13

Example Application  Five types of cups were selected.  The pattern is scanned pixel by pixel from left to right and from top to bottom using the 20 masks in order to calculate the frequency of each mask.  For example, the frequencies for Cup-l were calculated as follows: 112, 1423, 31, 27, 56, 55, 57, 56, 262, 267, 265, 261, 195, 196, 197, 201, 5, 5, 208, 218, Cup-l Automated Visual Inspection Using Inductive Learning Slide 14 Inspection of water glass cups:

Example Application (cont.) Set of examples for the five cups:  112, 1423, 31, 27, 56, 55, 57, 56, 262, 267, 265, 261, 195, 196, 197, 201, 5, 5, 208, 218, Cup-1  622, 840, 27, 39, 155, 154, 162, 158, 102, 104, 103, 101, 31, 29, 34, 36, 26, 28, 107, 111, Cup-2  230, 621, 37, 40, 22, 22, 22, 22, 116, 109, 109, 116, 18, 18, 19, 19, 45, 45, 18, 18, Cup-3  697,715, 2, 1, 10, 10, 10, 10, 91, 94, 91, 94, 10, 11, 9, 11, 4, 3, 1, 2, Cup-4  3739, 622, 72, 77, 557, 575, 560, 579, 144, 155, 154, 138, 521, 533, 523, 533, 542, 543, 110, 113, Cup-5 Automated Visual Inspection Using Inductive Learning Slide 15

Example Application (cont.) Rule 1  IF 112 =< Ml < 303 AND 1395 =< M2 < 1438 AND 30 =< M3 < 34 AND 25 =< M4 < 29 AND 39 =< M5 < 68 AND 179 =< MI6 < 207 AND 218 =< M20 < 230 THEN CLASS IS Cup-1 Rule 2  IF 494 =< M1 < 685 AND 836 =< M2 < 879 AND 26 =< M3 <30 AND 37 =< M4 < 41 AND 155 =< M5 < 184 AND 130 =< M8 < 160 AND 100 =< M9 < 109 THEN CLASS IS Cup-2 Rule3  IF 112 =< M1 < 303 AND 621 =< M2 < 664 AND 34 =< M3 < 38 AND 37 =< M4 < 41 AND 10 =< M5 < 39 AND 33 =< M17 < 62 AND 32 =< M18 < 61 THEN CLASS IS Cup- 3 Rule 4  IF 685 =< Ml < 876 AND 707 =< M2 < 750 AND 2 =< M3 < 6 AND 1 =< M4 < 5 AND 10=< M5 < 39 AND 91 =< M11 < 101 AND 2 =< M20 < 14 THEN CLASS IS Cup-4 Rule 5  IF 3550 =< M1 < 3741 AND 621 =< M2 < 664 AND 70 =< M3 < 74 AND 73 =< M4 < 77 AND 532 =< M5 < 561 AND 526 =< M17 < 555 AND 525 =< M18 < 554 THEN CLASS IS Cup-5 Automated Visual Inspection Using Inductive Learning Slide 16

Visual Inspection Applications Automated visual inspection has very large application areas:  Banknote recognition  Signature recognition  Fingerprint recognition  Number-plate recognition  Barcode recognition  Inspection of all parts, subassemblies, and finished products in mass production. Slide 17 Automated Visual Inspection Using Inductive Learning