Similar SKU Treads were getting mislabeled before shipment. Some SKUs have very similar tread patterns but vary in width by only 1/2 inch. With the product being black rubber, imaging the product can be quite difficult, and with 100s of different SKUs, pretraining all the SKUs is not feasible.
An auto-guide interacts with the PLC to get the current SKU and opens/closes it to the appropriate width, allowing the tread to be centered over a combination of 3D laser displacement cameras. The tread is imaged as it passes over an encoder / grooved conveyor rollers. The image is converted into a 2D image and contrast-enhanced.
A Setup Wizard allowed the local QA staff to easily train a new SKU in less than 10 minutes. Vision Scripts are stored on a central repository, allowing all three finishing lines to use the same files.
Track and Trace dashboard allows QA to review inspection results and replay images in case of a future customer complaint.
Before applying cement, the smooth side of the tread surface must be inspected to ensure the quality adhesion of the tread to a tire carcass. Although a human inspector can easily find defects like blisters and cracks, small defects can be a challenge, and other surface defect anomalies may require shifting positions to find variation in surface sheen.
Our team shadowed human inspectors while collecting images of the product, annotated images, and then an image-based deep learning model was developed to detect surface anomalies. Through further data collection, annotation, and model iterations, the system gained the ability to classify the surface anomalies into several categories. Traditional vision (rule-based) algorithms were added to established pass/fail criteria based on shape, size, and brightness for each classification, thus reducing the need for 100% human inspection.
A shrink wrap product sleeve is placed onto the bottle before entering into a steam tunnel. Occasionally the shrink wrap label is not in the correct position or gets snagged. This causes an unaesthetic package that is not suitable for retail. Each bottle must be manually inspected prior to filling. Shrink wrap colors are anything across the rainbow and many shades in between, including shades of white placed on a white bottle.
The bottle is side-belt transferred to an exit conveyor, and the bottle is inspected with a GigE camera. Using a combination of UV and white LED lighting, along with a specialty cut filter, the UV lighting causes a blue shift in the color of the product sleeve, while the color of the white bottle appears to be grey. Due to the reliability with traditional color segmentation working with greys, an in-house custom color segmentation routine was used to segment the exposed bottom of the bottle that is not covered by the product sleeve. This allowed the bottle to be profiled and measured, allowing for a quantitative pass/fail decision.
When customers are finished with a laser ink cartridge, they often send it back using the return box provided with new cartridges. These cartridges are returned to centralized recycling centers across the country by the truck loads. Used cartridges are removed from their packaging and thrown onto a conveyor belt for manual sorting. As volumes steadily increase, additional workforce is required to aid in the sorting. Since the recycling center is credited for each cartridge processed and specific cartridges have a higher number of recyclable components than others, an accurate count of models is highly desired.
PC-based vision system monitors the conveyor for cartridges as they enter into the field of view. At this point, it makes note of the cartridge's encoder location and its orientation. Using a blob tool, the system pre-classifies cartridges based on their shape using a high-end multi-core processor for parallelized pattern matching. Subsequent secondary pattern matching is performed in some instances to identify other key markers when subcategorization is required. A local register retains counts of each cartridge identified. Encoder location, orientation, and cartridge ID are passed to a spider robot to pick and place the cartridges into large bins. Multiple instances and orientations of well over one hundred cartridges were trained into the system.
The inside of a tire is manually inspected for surface defects and penetration by foreign objects. It can be physically awkward and challenging to see small objects manually, so the vision system handles that task.
For this pilot project, a camera on a linear slide and rotating mirror constrained to a 12" x 3" package were mounted on the end of the linear arm and joysticked into the center of the tire. Laser gauges provide feedback to the operator for their ideal placement. Interior is laser profiled to calculate optimal camera distance such that the entire field of view is within focus. The mirror is centered to the field of view, and the tire is rotated while the camera acquires a series of images. After each series of images, the mirror is adjusted to the next field of view, and the process is repeated until the interior is fully imaged. Images are then presented to the operator on a 34" curve monitor for manual inspection. The operator reviews the images and draws boxes around possible defects. Afterward, this information is archived with an identifier. At a later date, this data can be used to train a Deep Learning system.
QR code and human-readable text on the trading card would occasionally not match, either due to misconfiguration by the Operator or buggy remote software by the Vendor.
The system was deployed with surplus stock in less than two weeks, using 4VT's standard framework and a VisionPro Quickbuild script. Cards were fed onto a vacuum belt, triggered by a photoeye, and if there was no match, the system would blow the product off at the air gap between the vacuum belt and the accumulator belts at a rate of 10 cards per second.
Animal Health Sciences start-up was seeking a subject matter expert to develop a machine vision solution to determine the orientation of a baby chicken's head for the application of vaccine.
For this SBIR Phase 1 project, a custom lighting and optical solution were developed to acquire images of baby chicks as they passed under the camera. 1000 images were collected and annotated, 2/3rd of the images were then fed into Deep Learning algorithms. Detection performance was 97%, with an execution time of 19-22ms.
A Temp labor team assembled packaging for box displays for a Trading Card game. Occasionally cards or accessories were incorrect or missing.
The product enters a multicamera light tunnel and a QA engineer uses a Wizard to draw ROIs for each card and accessory. Cognex VisionPro's PatMax was used for its robust pattern matching algorithms. The system would stop the conveyor once it exited the light tunnel if there was a mismatch or no match, and turn on a stack light.
Seeds are collected and manually placed into a 96-well plate with gaps left open for control samples. These blocks enter into a machine that crushes the sends, which are then transferred over to another machine for genetic testing. The process of manually placing the seed in most, but not all wells of the test plate, is tedious and prone to mistakes. Also, it is important that one and only one seed is deposited in the appropriate well.
Seeds were dispensed into a small backlit bowl, A camera identified a single seed, with sufficient free space around it, so that a pick and place robot with a vacuum tip can move the seed from the bowl into one of the test wells of the 96-well plate.
A secondary camera, verified only one seed made it into the well of the plate. Certain seeds were prone to clog or gum up the surface of the vacuum tip and the secondary check verified compliance.
Motor components are assembled in stages, with the base attached to a carriage that moves through the assembly line. After the rotor is mounted in the stator the alignment needs to be checked.
Due to inconsistencies in the placement of the motor base on the carriage and the tightness of tolerances, a single top-down view and gauge approach were not viable. A stereo camera setup allowed for the end of the rotor and stator to be individually identified in 3D space. From there concentricity measurement and plane parallelism could be calculated.
Car interiors are often customized for their material, trim, and user features. Variations may be subtle and batch runs are small. An end-of-line inspection system is needed to ensure the product that was assembled matches the build ticket.
A camera with a bright field and dark field configuration was mounted to the end of an ABB robot. The operator loads an assembled car door interior onto an open nest and presses the Inspect bottom. The robot would move to 20 different locations on the backside of the car door interior, triggering the camera/lights, and then requested the operator to rotate the fixture. The robot then moved to 6 different locations, triggering the camera/lights for each.
A mock display is presented to the Operator for pass/fail of all zones. The operator had the ability to select a zone to see the image taken and individual pass/fail results for each Poka-Yoke item.
After the bearing raceway has been machined to a sized blank, the possibility exists that the surface may have underfills, pitting, or other blemishes. Such non-conformances need to be removed and repaired before the additional value is added to the product.
The product is placed into an encoder roller system via an overhead robotic system and is imaged with a line scan camera. If variations in the surface finish are detected, the product is moved to a secondary inspection area, other wise it is moved onto the next step in the process.
Prior to loading the tire for shipment, the highpoint is located and the sidewall of the tire is marked with a red dot, white dot, white P, white N or a sticker. Storage and material handling of the tire can cause the highpoint mark to get rubbed off or faded.
Cognex VisionPro was used to identify and verify the presence/absence of the various types of highpoint marks based on the product SKU.
Tires of similar sizes can be produced with different tread patterns. A color band is applied on the tread surface based on the product SKU to ensure the wrong product is not mixed in with the shipment.
Cognex VisionPro was used to blend the color band together (because of the gap between tread lugs) and a color match is performed to verify it matches with the current SKU.
A fuse block assembly for a car or truck has similar frames, but what goes into them can vary. A component can easily be missed, such as a fuse or diode or the wrong relay model installed.
A GUI was developed to allow an Engineer to train a new component (color, pattern matching, OCR) and add it to a component family. Once populated with components, an Engineer can create a product SKU and create a Region of Interests (ROIs) for each component location on the fuse block. The engineer can choose the allowed orientation of the component for the ROI.
With a fully configured product, an Operator can recall the SKU from a part list and batch inspect multiple fuse blocks. Any component that is missing, in the wrong orientation, pattern score is too low, or OCR is mismatched, then the ROI is highlighted in red. Components that pass are highlighted in green.
Key blanks for major retailers are plated, stamped and channels cut. Non-conformances can include poor plating, poor stamping, misaligned stamping, and misaligned channels.
The system was trained with the average composite image of 10 or more keys. Cognex PatInspect and additional image processing were used to determine the difference image, then blob analysis determined pass/fail result.
Opiate Lozenges are sonic welded onto a plastic stick, inspected, and transferred to a packaging tray. An embossed number on the lozenge indicates the dosage of the product. The wrong or unreadable embossed number is considered a non-conformance. A chipped or cracked lozenge is also considered a non-conformance.
Cognex VisionPro was used to pattern match the embossed number; too low scores were rejected. Image processing tools masked out the embossed number, and edge detection algorithms were used to identify cracks/chips.
The lot and date code are printed on a label, and the label is applied to the bottom of the stick holding the opiate lozenge. Missing label, missing print, or incorrect lot/date code is considered a non-conformance.
Cognex VisionPro was used for verifying label placement and OCV of lot and date codes.
From new to end of life, welding tips used to spot weld the housing to the band create varied results. From time to time, a weld would fail to fully engage and would produce a light or no weld.
5000 images were collected and annotated, then applied to a neural network-based descision engine (predecessor to modern deep learning algrothims). System performed well after new tips were broken in after 2000 cycles. (Average life cylce of tips 35k - 40k)
International mail brought in through airline processing centers are often in large sacks with a 4”x6” card attached to it. Depending on the country of origin, some of these cards are hand written and others are printed. The individual at the processing window needs to manually enter in all of the information on the mail tag into a data entry system before it can be thrown onto a conveyor belt to be processed. This creates a bottle neck in the work flow, since only so many transfer windows can be manned and managed to allow baggage/cargo trucks to be unloaded.
A portable prototype imager with a flat surface allowed the mail handlers to position the mail tag for image acquistion. At the start of a new load the mail handler would enter in the Airline and Flight number, along with the number of pieces. The mail handler would take an image of that tag and then the local Vision PC would attempt to read all of the text on the tag, then present the image along with data to the mail handler. Once approved, the data would be saved to a database. If tag was not readable, the image was sent to a remote location where it was presented to a person for manual entry. This allowed for reconcilation of a load and moving mail bags as quickly as the manual imaging process would allow.
Supplier for the machines that manufacture the Permanent Residency Cards needed to ensure that the text printed on the card matches what is contained in the database. It was also to ensure the person’s face was printed clearly, and verify the presence of several security features.
Standard lighting techniques and Optical Character Recognition routines where used to validate the text and data that was contained in the database. Image correlation was used to validate the person’s face printed on the cards. Multiple different lighting techniques and LED wavelengths were used with different image processing alghrothims to verify the presence of all security features.
Supplier for the machines that manufacture the Driver’s Licenses needed to ensure the text printed on the card matched what is contained in the database. Also needed to ensure the person’s face was printed clearly, and verify presence of the engraved birth date on the card.
Standard lighting techniques and Optical Character Recognition routines where used to validate the text and data that was contained in the database. Image correlation was used to validate the person’s face printed on the cards. Photometric stereo imaging technique with IR LED lights was used to extract the engraving and filter out the print on the driver’s license.
Dashboard displays manufactured by automotive suppliers are hand assembled with different Legend Tags (door open, oil temp, washer fluid low, etc.). Before the product is packaged and shipped for assembly into a vehicle, 100% inspection is required to prevent quality chargebacks.
PC-based vision system reads the 2D barcode on the cluster assembly and queries a database. The database provides information on what Legend tags are to be installed in each of the 10 locations. Using pre-trained patterns, the system verifies all 10 locations on the cluster and prompts the Operator of any mismatches.
Dashboard displays manufactured by automotive suppliers must have all the gauges (speedometer, tachometer, engine temp, etc.) calibrated before shipping.
The operator places the product into a test docking station, and the gauge is provided a signal for 3 different values. The PC-based vision system determines the angle of the needle on the gauge and provides the value back to the test docking station. The 3 data points are used to calibrate the gauge, and values are saved to the EPROM of the dashboard display.
Vinyl floor manufacturer was looking to increase the consistency of their product by detecting foreign color chips, shade breaks, and "eye poppers." Vinyl flooring is made by grinding up various color chips, heating them, and rolling them flat. Since broken tiles are often ground and put back in the mix, any tiles left over from the previous batch run could introduce undesirable variances.
A PC-based vision system segmented the image for the tile. During the training, the process scanned and identified all the grouped colored signatures, including the background average color—tiles with background shade breaks and non-conformances were rejected. After multiple consecutive rejects, an operator was alerted.
An Automotive manufacturer needed to verify the correct roof has been placed on the vehicle prior to robotic welding.
PC-based Vision system identified key features of 6 different possible roof configurations and verified a match based on the VIN barcode.
An Automotive manufacturer needed to verify the correct fender has been placed on the vehicle prior to robotic welding.
PC-based Vision system identified key features of 4 different possible fender configurations and verified a match based on the VIN barcode.
Tier 1 automotive supplier manufactures car frames. Mounts that are bent or incorrectly mounted on the frame need to be detected before additional value is added to the product.
PC-Based vision system located and profiled the mount that was welded to the frame. An alert was raised if the location or deformation was outside of tolerance by more than 1/4 inch.
Molded fenders made from thermoplastic olefins (TPO) require an adhesion promoter (Adpro) prior to the application of automotive paint.
A PC-Based vision system projected a laser line onto the TPO fender. A raw fender would create a diffraction pattern when reflecting the laser line, whereas the application of Adpro creates a diffuse reflection. Image processing analysis detected the signature difference between the two, and the system would stop the line if a non-conformance were detected.
Fenders are welded to a truck bed by welding robots. Misaligned fenders need to be detected and corrected before the truck bed proceeds to the next step in the process.
The PC-Based vision system used structured light to detect and measure fender location on the truck bed. Rejects were flagged, and an alert was raised to the Operator.
Metal truck panels after forming enter into a rust inhibitor dip tank. Panels that are not properly secured can "float" off the carriage into the dip tank. Subsequence carriage entering into the dip tank can have their panels damaged.
A PC-based vision system was used to verify all the locking pins holding panels to the carriage were in the correct position. If an improper position locking pin was detected, the line was stopped, and an alarm was raised.