Avnet simplifies visual AI application development

With the rapid development of artificial intelligence (AI) applications, visual AI has become the main battlefield for various technology companies. Based on machine learning, visual AI devices at the edge of the network can complete various intelligent visual tasks such as object detection, face recognition, image analysis, etc. based on AI reasoning, bringing a new experience to users.

However, if you are in the "battlefield" of visual AI, you will deeply understand that this is a difficult "battle". The main challenges come from the following aspects:

• Completing AI reasoning in edge devices requires sufficient computing power while also meeting many constraints such as cost, volume, power consumption, and security. The design of products and solutions is not easy.

• Although the total size of the visual AI market is not small, it is seriously fragmented. According to incomplete statistics, there are more than 70 related visual use cases on the market, but they cannot be dealt with by a unified solution.

• The rapid evolution of AI technology is also a worry. The ever-changing algorithms, sensor technology and visual processing processes force developers to keep up and continuously optimize their designs.

Therefore, if you want to win in such competition, your company must not only have the courage to fight, but also have higher "wisdom" to find more efficient technical paths, establish its own technological advantages, and widen the gap with others.

Adaptive computing platform

Visual AI applications in edge devices generally include AI reasoning as well as non-AI pre-processing and post-processing functions, and all of these functions require corresponding high-performance computing power to support them. If a chip can be designed specifically for a specific visual AI application, it is of course the best acceleration solution from a performance perspective. However, the limitations of such a fixed chip solution are also very obvious: first, the R&D cost and time cost of dedicated chips will be very high; and this is obviously an irreconcilable contradiction with the fragmented market and rapidly iterating technology of visual AI applications.

In order to solve this problem, in terms of hardware, the use of FPGA SoCs that integrate programmable logic (PL) and Arm embedded processing systems (PS) is a good choice, such as Xilinx's Zynq UltraScale+ MPSoC. In this way, developers can use one device to meet the computing tasks of the entire process of visual AI processing; using the PL subsystem, developers can achieve the best deep learning processing unit, video processing and scalable sensor fusion based on specific use cases and based on the latest AI algorithms and processing flows, and find an optimal balance between high performance and flexibility.

Figure 1: Zynq UltraScale+ MPSoC system block diagram

In addition to hardware, a set of supporting software is also essential. Xilinx provides developers with a complete tool chain (as shown in Figure 2). No matter what level of developer you are, there are software tools that match the corresponding design path available for you to use.

Figure 2: Xilinx software development tools

The combination of these software and hardware forms a unique "adaptive computing platform". For systems that require both flexibility and efficiency, using an adaptive computing platform based on FPGA SoC for development is undoubtedly a wise choice.

4K Smart Camera Development Platform

After selecting a basic development architecture such as the adaptive computing platform, it is not all done. Developers will still face challenges in specific solutions and product development.

Generally speaking, the process of visual AI application development is as follows: first select the chip, build a prototype, and preliminarily verify whether the chip is compatible with the required AI model; then design the PCB and integrate the system, and optimize the software and hardware on this system platform that is closer to actual commercial use; finally, put it into mass production after finalization of the test.

However, in actual development, visual AI applications are diverse and multi-layered. If each application development starts with chip-level design and then goes through a complex system integration process, this requires a long R&D cycle and the participation of a complete hardware, software, and PCB development team, which invisibly raises the threshold for R&D.

Is it possible to skip the early chip-level development and PCB design for the development of visual AI applications and start from a higher-level, more complete product platform, or even directly use this development platform as a mature product to simplify the entire development process? Avnet's 4K smart camera development platform can meet your requirements.

Features of the Smart Camera Development Platform include:

• Equipped with 13-megapixel image sensor and high-performance image signal processing chip

• Xilinx MPSOC’s powerful computing power can support the deployment of large-scale neural network algorithms

• Vitis AI development toolkit provides a mature and rich AI model library

• VCU hard core embedded in MPSOC supports ultra-low latency AVC/HEVC codec

• The low-latency face detection reference engineering source code based on this platform is fully open to users

• Productized hardware structure can quickly transform user designs into final products

Figure 3: Smart camera development platform

We can see that this smart camera development platform is based on Xilinx's complete hardware and software adaptive computing platform, and is supported by a complete design ecosystem, which is conducive to accelerating the development of the final application... Application development, which was daunting in the past, has now become within reach.

During use, developers can either use it as a mature smart camera directly in system solutions, or take advantage of its powerful performance and flexible scalability to use it as an "AI-box" development platform to explore more possibilities of visual AI applications.

Figure 4: Face detection camera solution based on smart camera development platform

Professional technical services

With the support of adaptive software and hardware platforms and the selection of development tools such as the smart camera development platform, if developers still want to make visual AI application development more "easy", then Avnet has a good trick!

As a global technology distribution partner of Xilinx and a "powerful company" with many years of experience in the field of visual AI, Avnet can provide customers with a full range of application development technical support. The services include (but are not limited to): providing detailed Vitis AI application cases based on the smart camera development platform; supporting the transplantation of AI neural network models on customer boards; supporting the production and hardware customization needs of smart cameras; supporting the development of customized smart cameras based on Xilinx's new product Kria core board... With such technical support, customers can focus more on their own technical expertise or exploring market needs.

In short, to achieve the goal of "simplifying visual AI application development", we need an efficient platform, convenient tools, and professional services. Once you master the tricks, you can find quick solutions to even the most difficult design problems.