Vision AI Based Defect Detection on AM62A

Demo Background

Defect detection is a crucial part of quality assurance in the manufacturing process. This demo uses AM62A to run a vision based artificial intelligence model for defect detection for manufacturing applications. The model tests the produced units as they move on a conveyor belt, to recognize the accepted and the defected units.

An object tracker is developed for this demo to provide accurate coordinates of the units for sorting and filtering. A live video is displayed on the screen. The units are marked on the screen using green boxes for good (accepted) units while defected units are marked with boxes with different shades of red to distinguish the types of defects. The screen also includes a graphical dashboard showing live statistics about total products, defect percentage, production rate, and a histogram of the types of defect.

The code base and details of how to run the demo are available on TI marketplace as a Github repo: https://github.com/TexasInstruments/edgeai-gst-apps-defect-detection.

How to get started

Following are the steps to run the demo:

1. Get an AM62A starter kit EVM
2. Download the Edge AI Linux SDK
3. Load the Edge AI Linux SDK via an SD card using the quick start guide
4. Log into the EVM through a network connection
5. Clone the git repo for this demo onto the EVM (or copy all files to the SD card)
6. Run the ./setup-defect-detection.sh script. This downloads required pre-train model artificats and a pre-recorded test video to the EVM.
7. Open the apps_python directory.
8. Run the demo using ./app_edgeai.py ../configs/defect_detection_test_video.yaml. This command uses a pre-recorded video as input. The Github rep contains detailed instrucitnos about using a camera feed as input.

Deep Learning Accelerator vs ARM cores

The AM62A SoC is equipped with a Deep Learning Accelerator (C7x-MMA) with up to 2 TOPs at 1 GHz. The deep learning inference is offloaded to this accelerator. The defect detection application showed a frame rate of 30 FPS. This rate is limited by the performance of the camera used in the application which does not exceed 30 FPS. For this reason, only 22 % of the deep learning accelerator C7x/MMA is utilized. With this low utilization, the frame rate can scale up to +130 FPS if a faster camera was used. This high frame rate could not have been achieved using only ARM cores.

The video show side by side screen recording of the Defect Detection application when inference is executed on the C7x-MMA DLA (left side of the video) and when inference is executed on ARM A53 cores (right side of the video).

• The same physical setup is used in both experiment including the rotating speed of the table and the number of samples on the table.
• The C7x-MMA side shows great performance. The application is able to detect, classify, and annotate all objects on time with no delay.
• The ARM inference side shows very slow performance, most of the objects are not detected on time or never detected. The inference could barely work on 2 FPS.

This comparison shows that it is not feasible to use ARM cores for machine vision applications, such as defect detection, which require high frame rate and that the C7xMMA DLA is necessary to provide the required frame rate.

Additional Resources for the Defect Detection Demo

General Edge AI and AM62A resources

Please find the following resources related to the AM62A and TI Edge AI.