Inspection on the Surface
There are two types of lighting: diffuse and structured. When we talk about diffuse lighting, we really mean uniform light intensity from all directions. This really isn’t possible unless you intend to perform your inspection within an integrating sphere. In a manufacturing environment such a design is not feasible.
Rather we use something that can approximate the diffuse lighting environment from part of spherical surface or more likely a cylindrical surface. Phoenix Imaging manufactures a tunnel—o r 2040 Series—lighting system that provides a diffuse lighting environment for illumination of single plane object. The tunnel is sized to match the object under inspection.
The size is roughly three times the object in the X & Y dimensions (assuming that the object lies in the X-Y plane). The diffuse lighting system performs adequately for the isolation of the surface porosity on machined metal components. The porosity can be thought of as a spherical void which has been cut through during the machining operation. The majority of the porosity surface does not coincide with the plane of the machined surface and therefore does not reflect light back in the same direction of the machined surface, thus appearing dark compared to the lighter machined surface. This would appear to be a fairly simple inspection process of searching for dark blobs on a light surface.
However, the machining operation tends to produce subtle machining features that manifest as different textures across the machined surface of the component. In some places these textures may actually appear darker than the porosity that we are searching for. Simple thresholding (identification of pixels that are lighter or darker than and specific limit), usually does not produce consistent inspection results in this situation and more advanced image processing techniques are required to produce the desired result.
One of the more effective vision algorithms for surface inspection is the use of grayscale morphological operators to perform localized grayscale normalization. This technique compares each pixel in the image to its surrounding neighborhood and performs a combination of openings and closings to generate a relative contrast image. This relative contrast image provides a consistent grayscale difference between the porosity and machined surface everywhere in the image, even with some variation in the lighting uniformity. However, this is a fairly intensive set of image processing operations and requires sufficient processing power to perform at typical production rates.
A fairly high speed processor with a enough memory is required to perform these operations when using small images and the system may require several processors to handle very large image formats. Fortunately computer processing speeds are increasing along with faster memory and some image processing companies are taking advantage of the multiple core Graphical Processing Units (GPU) to reduce algorithm processing times. These new GPU systems are reminiscent of the massively parallel processors known as Single Instruction Multiple Data (SIMD) architecture that was popular in the 1980’s. These superfast image processors would implement 128, 256, 512 or even 1024 processors operating in parallel to process images extremely fast. They were ideal systems for computationally intensive surface inspection applications with one exception, they were very expensive. The major difference between the old SIMD systems and the new GPU systems is that the price is now about 10% of the former with similar performance. The multi-core GPU units are provided in a graphic board PCI-e format and must be used in the standard image processor configuration rather than a “smart camera” configuration.
Even the more standard Intel Central Processing Unit (CPU) is evolving into a multiple core configuration. The current Intel i7 CPU has 6 cores operating above 3.4 GHz and in the near future will have even more. The advantage of the multiple core processors is that the software can be written to perform different task on each core while sharing information in memory. The real benefit of a multiple core image processor is that each core can be configured to handle a separate sensor and each can operate independently of each other. This allows the customer to purchase a single image processing system and then add sensors as required at a much reduced cost. For example, if the surface inspection application requires both bright and dark field inspection techniques to isolate the defects both can be performed simultaneously using two different sensors and optical setups. Because the images are located within the same memory space the correlations of defect locations can be performed rapidly without have to transfer images between devices.