Until now, embedded vision applications have mostly used very simple sensor modules, which were mainly developed for consumer electronics applications and not designed for industrial applications. Developers of embedded systems have therefore had to compromise on image quality and camera performance in favor of size, power consumption and price. However, the advantages of embedded solutions, such as lower costs, low energy consumption and compact design, also make a switch from PC-based to embedded solutions in the industrial sector interesting and require new approaches.

Image quality versus computing capacity

When integrating vision into embedded systems, developers currently often rely on CMOS sensor modules to capture images. The functionality of these devices is limited to the connection of the image sensor to the host - usually an embedded board. When integrated into smartphones, tablets and laptops, these small and inexpensive modules deliver acceptable image quality.

However, this is precisely the disadvantage of this camera option for embedded vision: such sensor boards have very low image processing capacities in the camera and only allow minimal pre-processing. The embedded community has come to terms with shifting important image processing tasks to the CPUs (Central Processing Units) or dedicated ISPs (Image Signal Processors) such as GPUs (Graphics Processing Units) or MPUs (Media Processing Units) on the embedded board. Both image corrections to improve image quality and application-specific image processing run on the embedded board. This reduces the available performance and computing capacity of the embedded board's central processors, meaning that less computing power is available for the important image processing algorithms. To compensate for this shortcoming or to maintain the desired performance, increasingly powerful and therefore more expensive boards are required.

Requirements for image processing

Today's preferred CMOS sensor modules were originally developed for cell phones, but are also suitable for other applications due to their attractive price, small size and low power consumption. Most modules have a number of basic algorithms and automatic image control functions for image processing, such as automatic exposure, white balance or black level calibration. Above all, however, they use simple algorithms to improve image quality, for example by optimizing image sharpness, lens and defect pixel corrections or noise suppression. These are standard functionalities in modern CMOS sensor modules that are configured by designers so that the embedded board delivers acceptable images.

For many applications and tasks, however, the image quality that can be achieved is no longer sufficient. The demand for more powerful image processing systems at the level of machine vision cameras is increasing rapidly. Machine vision consists of three stages of image processing: Pre-processing, advanced image processing and application-specific downstream processing. In each step, different algorithms are used to process and analyze the image.

The first step after capturing a RAW image that has not yet been modified by algorithms is image pre-processing. This image correction and optimization is the sum of the processing tasks performed, which convert the raw image data supplied by the sensor into an image whose quality meets the requirements of the respective system application. Such image pre-processing functions can include, for example, algorithms for pixel correction, white balance, amplification or noise reduction, which optimize the image for the task at hand for the image processing software. After these image optimization steps, advanced processing - sometimes just called image processing - enables more complex enhancements to facilitate application-specific image analysis. Examples of such advanced processing include sharpness or color correction using look-up tables.

The final step is post-processing, which performs a wide range of tasks depending on the system specifications, such as identifying faces in images, locating and reading license plates in traffic monitoring applications or checking the quality of objects in an industrial inspection application. These tasks are performed by complex, application-specific software algorithms on the embedded board's main processor.

More efficiency required

When using a sensor module, both the extended and the application-specific image processing takes place on the embedded board. This means that the embedded board takes over a large part of the tasks and the necessary computing power, as the sensor module cannot do this. This puts a strain on the performance of the embedded board: not only does it become slower, but there is also less performance capacity available for possibly more important other tasks.

If no demanding algorithms have to be used, a simple board may be able to cope with this without any major loss of performance. However, if the requirements for image processing algorithms and image quality become more demanding, an additional co-processor or a dedicated ISP on the embedded board is required - because when it comes to more advanced algorithms and image processing such as special filter, pixel and signal processing, CPUs reach their limits early on.

One alternative for overcoming this capacity bottleneck is to choose a more powerful embedded board. However, when developing cost-sensitive embedded systems, these more expensive boards are often not an acceptable option, which is why it is not uncommon for image quality to be compromised.

Shifting image processing to the camera module

Another way to meet image processing requirements without having to resort to more expensive embedded boards is to shift the image processing tasks to the camera. Following this approach, Allied Vision has developed a new type of camera module that combines the advantages of an embedded sensor module with the performance of industrial image processing cameras. The camera series is based on the 'Alvium' technology developed by Allied Vision, an Application-Specific Integrated Circuit (ASIC) with integrated ISP and a proprietary image processing library that contains the company's accumulated machine vision know-how.

The core of the 'Alvium' series is the integrated, robust ASIC, which is perfectly suited for embedded vision applications thanks to its low power consumption.

© Allied Vision Technologies

A major advantage of the concept of performing more image processing in the camera instead of on the embedded board is the reduction in processor and co-processor load. This frees up resources for potentially other tasks.

This technology leads to various other advantages of the 'Alvium' camera modules, which are all based on the proprietary ASIC chip. The 'Alvium 1500' series offers a limited range of functions that is specifically geared towards the needs of the embedded vision community, while the 'Alvium 1800' series offers an extended range of functions to meet all requirements for simple machine vision and industrial embedded vision applications.

In contrast to Field Programmable Gate Arrays (FPGAs), which are often used in machine vision cameras, the Alvium ASIC has a much higher integration density, which makes it very energy efficient. It also has intelligent power management, which means that only the functionality currently in use requires power and the rest is switched off. This results in a lower temperature rise and the camera does not run the risk of becoming too hot.

The camera series incorporates the experience gained from countless industrial image processing projects; this is reflected in the industrial suitability of the camera module for embedded vision. The Alvium camera platform is based on an ultra-compact single board design with all electronic components on both sides of a 27 mm x 27 mm board, which is available as a bare board module. This ensures small size and low weight and is also the most robust solution for applications exposed to shock and vibration.

The 'Alvium' camera bridges the gap between machine vision and embedded systems. Thanks to increasingly powerful boards, the switch from PC-based to embedded systems is also a good option for demanding industrial computer vision applications.

Author:
Gion-Pitschen Gross is Product Manager at Allied Vision Technologies in Ahrensburg.