How to Deploy Deep Learning Object Detection on Intel® Hardware

Visual computing has used OpenCV algorithms to detect objects for decades. Deep learning inference takes computer vision to entirely new levels of sophistication with support for poor lighting, off-angled shots, and subtle flaws.
 

What exactly is deep learning object detection?

Deep learning object detection combines two computer vision tasks: localization and classification. In localization, the model identifies objects in an image and draws a bounding box around them. In classification, the model compares the objects with a known list of classes and predicts what each object is. These predictions are usually output as labels with a name and a confidence rating.
 

What can you do with object detection?

Deep learning object detection is in near-constant use across business and industry. Search engines use it for image-based search. Social media sites use it to identify individuals in photos. On the edge, individuals, businesses, and industries use object detection for things like …
 

• Animated emojis for social media
• Automated checkout and loss prevention at grocery stores
• Defect detection and quality control in manufacturing
• Smart medical imaging and devices
• Predictive maintenance on machinery and factory lines
• Worker safety
• Autonomous controls for robots in warehouses and assembly lines
   

How does deep learning object detection work on edge devices?

In most cases on the edge, object detection needs to occur on the device in near-real time. For example, an object-detecting robot on a factory line can’t wait for data to travel to the cloud and back to receive its operating instructions.

Object detection in more extreme applications—think drones, satellites, Mars Rovers—requires deep learning models and inference engines that can run on less power and lower-performance processors with minimal memory, storage, and network bandwidth.

These are major challenges, but even extremely low-power processors, like the SoCs inside smart cameras, can run AI workloads.

Deep learning object detection use-case example: Industrial welding robot

Modern automobiles have thousands of welds in them, almost all of which are done by robots. Let’s say we want our robots to visually inspect their welds as they make them.

Giving our robots smart eyes will require high-definition cameras and a deep learning model that can analyze things like weld shape, consistency, and porosity—skills that human welders learn through years of training. On top of that, the robot has to capture the video, analyze the weld, and act on what it finds in near-real time.

Here’s how to put together a smart welding system that can perform at the edge practically and economically.

1. Choose, or create, the right deep learning model
   When you have a single, focused task, you only need to recognize discrete objects. Our robot needs to identify pinholes and irregular shapes, not items at a cash register like apples, oranges, and barcodes. Removing unnecessary objects from the model makes it smaller, and smaller means faster runtimes in deployment.
    
2. Optimize for the deployment hardware
   Our robot might have an x86 CPU with an integrated GPU or an ARM CPU with a video coprocessor or be controlled by an edge server with an AI accelerator—or it might be a mix of all these hardware types. To maximize inference performance, you have to convert the deep learning model into executable files for each hardware type and load balance across the different architectures at runtime.
    
3. Balance performance and accuracy
   Switching processing formats—from FP32 to FP16 or int8—can dramatically cut inference-processing time. You do lose some accuracy, but the penalties are usually very small compared to the performance gains.
    
4. Add it to the equipment and computing systems you already have
   If we’re lucky, our robot control system runs on open standard, off-the-shelf hardware and open, or modifiable, control software. If it does, all we need to do is install the camera systems and upload new software. If our robot runs on a closed system, we’ll need to talk to the manufacturer.

Sound like fantasy? It’s not. We helped Audi build a proof of concept for robotic weld inspection that increased their inspection rate from one car a day to every weld their robots make. The details are in our case study for Audi’s deep learning inspection system. 

Audi is just one industrial client that is using Intel® hardware, development tools, and software to put object detection and other deep learning models to work on the edge. Here are some more case studies about AI deployments at the edge.

Vulcan AI: Deep learning deployments for worker safety 

DeepSight: Deep learning defect detection for tire manufacturing 

ADLINK: Visual inspection for contact lens manufacturing

ExxonMobil: Automated wellhead controls and analytics
 

Want to learn how to deploy deep learning AI on Intel? Check out the Intel® Edge AI Developer Certification Program

The Intel Edge AI Developer Certification Program is a four-part—and growing—course on the basics of building and deploying deep learning applications on Intel hardware. The four courses cover:
 

• Edge AI fundamentals
• Object classification and retail loss prevention
• Object detection and industrial worker safety
• Style transfers and smart cities data augmentation

Each course features an instructional video, exercises with a sample application, and a final assessment. You can take them at your own pace, and the materials—including an account on Intel® DevCloud for the Edge—are free for your use. You can get officially certified for USD 99 a year. Certification comes with digital badges and credentials that you can use in your resume, portfolio, and social feeds.

New courses will be coming online so that you can continue building your AI deployment skills and learning increasingly complex material. If you’re interested in putting AI to work, check it out.
 

Learn deep learning deployment with the
Intel Edge AI Developer Certification Program.

Notices and Disclaimers

Intel is committed to respecting human rights and avoiding complicity in human rights abuses. See Intel’s Global Human Rights Principles. Intel® products and software are intended only to be used in applications that do not cause or contribute to a violation of an internationally recognized human right.

Intel® technologies may require enabled hardware, software, or service activation.

No product or component can be absolutely secure. 

Your costs and results may vary.

© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.