The processing needs of computer vision: a growing trend

It is a cutting-edge technology that allows obtaining, processing and analysis of any type of information obtained through digital images. Stephen Turnbull, director of Vertical Markets for AMD's Integrated Solutions division, explains in this article how the teams that analyze the images can do so in increasingly more detail and, therefore, the potential of this discipline increases, as do the processing needs.

Tell your friends and family about computer vision and they'll probably look at you funny., and they may immediately start chatting about a movie they saw in which robots became aware of themselves and became dangerous. Fortunately, reality is not so sinister. Perhaps the industry should consider naming this technology segment that is easier for users., as this cutting-edge field has enormous and positive potential for embedded applications.

In its essence, Computer vision simply consists of taking advantage of the information available in an image to make a decision about what to do, next, with the object of the image.

A simple pass/no-go when examining a product on the assembly line or before shipping is one of the simplest examples.. Inspection of a printed circuit board (PCB) It's a common use case., where the image of a correctly made master board can be easily and quickly compared to the PCBs being produced, as they move from an automated pick-and-place system to the next stage. This is a valuable step to ensure quality and helps reduce scrap items., and that the human eye and brain could not constantly repeat hundreds or even thousands of times per day.

Seeing is believing in integrated computer vision

As the resolution of image capture systems increases, the potential of computer vision does too, because the level of detail available to perform an evaluation increases at a corresponding rate. Increasingly smaller subsets of visual information can be evaluated by comparing them to a master template., increasing the load on the system processor to process the data and quickly provide a decision on what to do next steps (passes/does not pass, retain, start from the beginning, etc.).

The classification of vegetables is a case where their categorization as suitable/unsuitable is not optimal., since the standards are different from one country to another and the quality of the product varies throughout a season. To be able to minimize waste for the producer and still maintain adequate quality for the customer, more optimal algorithms are needed for quality classification, an almost impossible task for the human eye and brain.

A company that is using this application is the Danish Qtechnology. The company offers smart cameras for the classification of vegetables, with production volumes of up to 25 tons per hour, which requires analyzing more than 250.000 products between around 500.000 images. A 6,2 MB for each image, This particular case requires the analysis of more than 2,5 Terabytes of image data every hour per machine, a colossal amount of information to process. With this amount of data, more than 6 hours of transfer time on a Gigabit Ethernet connection.

Solving this with simpler algorithms would require multiple stages and cameras., lighting on the machine, more real estate in factories, etc. The alternative is to apply extensive processing capacity, either as a centralized processing unit via broadband connections or distributed processing with smart cameras, processing real-time data directly in the camera with only the results per product offered by the final mechanical sorting system.

To cope with different image capture technologies, Qtechnology uses interchangeable heads with different sets of sensors that go with smart camera systems. Your head of hyperspectral images, For example, Enables non-destructive detection of food quality and safety.

In standard vision systems, Food quality and safety are generally defined by external physical attributes such as texture and color. Hyperspectral imaging is giving the food industry the opportunity to include new attributes in quality and safety assessment, such as chemical and biological attributes to determine sugar, the fat, humidity and bacteria count in products.

In hyperspectral images, Three three-dimensional image cubes with spatial and spectral information are obtained from each pixel.. A greater number of spectral features offers better discrimination of attributes and allows grading based on more features.. Image cubes include intensity (reflected or transmitted light) of each pixel for all acquired wavelengths of light, resulting in each image cube containing a mass of information. This amount of data represents an exponential increase in the computational challenge to extract qualitative and quantitative results for product classification in real time..

The application of heterogeneous computing

Support for these processing demands, today and in the future, requires high-performance and scalable processing.

Qtechnology uses an accelerated processing unit (APU) in its platforms for smart cameras that combines the GPU and CPU on the same chip, allowing the system to offload processing of pixel data in vision applications to the GPU without high latency bus transactions between processing components.

This allows the CPU to serve other interrupts with lower latency, helping to improve real-time performance of the entire system and address the increasing processing demands of modern vision systems.

Pairing a different processing engine on a single wafer or system to apply the right processing power to the problem is at the heart of heterogeneous computing..

The Heterogeneous System Architecture Foundation (HSA) was established in 2012 specifically to help the industry define open specifications for processors and systems that take advantage of all available processing elements to improve processing efficiency.

The GPU is a massively parallel engine that can apply the same instructions across large data sets. (in this case, pixels) at the same time; That's exactly what it takes for you to run a 3D game on your favorite gaming console or PC..

Coincidentally, this is also exactly what is needed for computer vision. Additional performance can be increased by pairing the APU with an external discrete GPU, in a Mobile PCI Express Module format (MXM), allowing businesses to add additional GPU processing resources to support even more intensive vision tasks when needed.

Software is a critical part of the equation. Con HSA, The entire processing platform can be governed by a standard Linux kernel, requiring only modest development support with each new kernel release. The Yocto Project, an open source collaborative project, provides templates, tools and methods to help users create custom Linux-based systems for embedded products.

Huge ecosystem support for x86 allows companies to leverage third-party and open source image processing libraries, like OpenCV, Mathworks Matlab y Halcon. Debugging tools, latency analyzers and profilers (perf, ftrace) are also widely available.

Stephen Turnbull
Director of Vertical Markets of the Integrated Solutions division of AMD.