By using multiple cameras integrated into a single system, comprehensive inspection of small parts can be achieved.
While the quality of small parts such as washers, screws, and rubber pads can be inspected manually, this is a tedious, time-consuming, and repetitive task. Moreover, such inspections are prone to errors due to human eye fatigue, and manual inspections cannot be completed at high speeds. Therefore, such inspection tasks are better suited to be performed by machine vision systems.
In addition to performing such inspection tasks faster, machine vision systems can image each part from multiple angles, ensuring that all parts of the part are manufactured correctly. To achieve this, machine vision systems require the deployment of multiple imaging stations, which typically employ different lighting sources, camera systems, and lens products.
UTPVision has developed a rotary table-based inspection system to perform this task (see Figure 1). In operation, the part is loaded onto the rotary table via a vibrating bowl, and then rotates on the table past multiple imaging stations to inspect for top defects, bottom defects, internal defects, and other defects. The system uses multiple vision inspection stations to perform comprehensive inspections of small parts, such as washers, screws, and rubber pads.
Figure 1: To develop a machine vision system for inspecting small parts such as washers, screws, and rubber pads, UTPVision used a rotary table with three vision stations.
Size and surface
After the vibrating bowl feeder (VBF) loads a small part onto the system's rotary table, the system rotates, feeding the part into the imaging station (see Figure 2). As each part rotates, a Keyence laser displacement sensor measures the part's height, and after a set time delay, triggers the camera at the first imaging inspection station. Here, the part's dimensions are checked.
Figure 2: During operation, the part is loaded onto a rotary table via a vibrating bowl and then passes through multiple imaging inspection stations on the rotary table to inspect its (a) dimensions, (b) top defects, (c) bottom defects, and (d) internal and peripheral defects.
At the first imaging inspection station, a Manta G-504 GigE camera from AlliedVision was mounted above the rotary table. This camera uses a 2/3-inch 2452×2056 CCD sensor and is equipped with a telecentric lens from OptoEngineering. The camera captures color images of the parts and transmits them to the system's host PC via its GigE interface.
Next, the part is moved by a rotary table to a second imaging inspection station, where a camera captures an image of the part's upper surface. This second imaging station also houses a Manta G-504GigE camera equipped with a Computar 35mm lens. For accurate imaging of curved surfaces (often highly reflective mirrors), a dome lamp is used as the illumination source, positioned close to and above the part being imaged, to clearly reveal its angles, textures, or morphological features.
Different colors
Because different parts being inspected may reflect and/or absorb different wavelengths, the custom-designed dome lights used in this system include both red and green LEDs. Once a part enters the imaging inspection station, it is immediately illuminated by diffused red and green light.
In this way, illuminating the part with red light will cause the red-colored surface features to reflect, while the corresponding green-colored surface features will appear darker. Similarly, illuminating the part with green light will cause the green-colored surface features to reflect, while the red-colored surface features will appear darker. After the part is illuminated, an image is taken under each of the two different light sources. These two images are then combined and transmitted to the system's host PC via the camera's GigE interface.
In addition, the system also deploys a similar imaging station below the rotating disk to capture images of the surface features on the bottom of the part, which are then transmitted to the host PC.
Internal parts
In addition to capturing the height information and top and bottom surface features of a part, it is usually necessary to capture the features of the outer and inner surfaces of the part. For example, in the case of inspecting a rubber ring, it is necessary to analyze the parting line on the outer surface of the part, as well as any defects that may appear on the inner surface of the part.
While using dedicated outer-ring lenses and catadioptric outer-ring lenses can reduce the number of cameras and lenses required for this task, in some cases, the resolution achievable by such lenses is insufficient to meet the application's detail requirements. In such cases, a multi-camera/lens solution is necessary.
To perform the inspection of the inner and outer surfaces of parts using a rotary table-based inspection system, UTPVision developed a dedicated imaging system that utilizes 12 cameras. When a part rotates under the imaging system, six cameras are triggered to image the part's outer surface; these six cameras are all equipped with Sony 1600×1200 CCD sensors and mounted at a 60° angle to the horizontal. Simultaneously, another six cameras, also equipped with Sony 1600×1200 CCD sensors, are triggered to capture images of the part's inner surface; these six cameras are mounted at a 45° angle to the vertical. The images captured by these 12 cameras are then transmitted to the system's host PC via a GigE interface.
When images are captured by the system, they are displayed on the graphical user interface of a flat panel monitor (see Figure 3). This interface, written in GTK+, allows users to clearly see (clockwise from top to bottom) the dimensions, surface features, outer edge features, and inner surface features of the part. Therefore, the operator can see the images captured at each imaging station while the system processes the images.
Figure 3: In the design of this system, UTPVision utilizes the Retina library to analyze images using supervised learning. With this user graphical interface, the operator can see (from top to bottom) the dimensions, surface characteristics, outer edges, and inner surfaces of the part.
Supervised learning
Instead of using standard image processing algorithms such as feature analysis to detect specific features and color defects in each part, this system employs Retina, an artificial intelligence software based on a C/C++ library, developed by Squeezebrains, a division of UTPVision. Unlike unsupervised or semi-supervised learning algorithms, Retina uses proprietary algorithms to analyze images, requiring operators to train the system with a set of images. Therefore, parameters need to be configured; these parameters are the images used for training. In this supervised learning system, the algorithm uses input variables (in this case, images) and output variables (pass or fail) to learn a mapping function from input to output.
In this system, Retina software is used to process all images captured from various machine vision stations. This eliminates any image preprocessing stages, such as image thresholding to preserve all information about the image.
The number of parts required to train the system depends on the variability of the parts being inspected. For simple parts, the training process requires approximately five parts; for more complex parts, the number of parts needed may be greater. Typically, approximately 100 images of an object are captured first. Then, images of acceptable and unacceptable parts are imaged, and the operator inputs these images into the system to train it. Once training is complete, other parts are presented to the system, and the system queries the operator about the condition of the parts to reinforce supervised machine learning.
After being sorted, the parts rotate around a rotary table. To sort the parts according to defects, a PLC connected to the system host triggers Festo's air-blowing valves, placing the parts into different bins. By configuring the system in different ways, parts can be sorted as "qualified" or "unqualified," or they can be placed into different bins based on the type of defect detected. In many cases, defective parts can be recycled or remanufactured after sorting.
