Vision inspection adaptation usually includes equipment and operating system confirmation, camera access, algorithm running, Web or local interface, workstation I/O, deployment maintenance, and on-site acceptance. Being able to start the software is only a part of the adaptation; only when the input, algorithm, results, interface and version management form a closed loop can the device-side adaptation have engineering usability.
The current assembly visual inspection software and algorithm applications are already running on Huawei Ascend Atlas 200I DK A2. The operation of the platform indicates that the software, cameras, algorithm calls and result links have been connected; the production acceptance of the workstation still needs to be completed in combination with specific products, independent samples, rhythm and on-site interfaces.
Step 1: Confirm Atlas hardware and software environment
Before adaptation, the device model, processor, operating system, driver, firmware, CANN or other runtime version should be recorded, and confirm whether the storage, memory, network and peripheral interfaces meet the project requirements.
The official specifications of Atlas 200I DK A2 include Ascend 310 series processor, 8 TOPS INT8, 4 TFLOPS FP16, 4 GB memory, USB, MIPI and Gigabit network interfaces, and supports Ubuntu 22.04 and openEuler 22.03. The nominal specifications of the equipment are used for platform selection, and the actual detection performance still needs to be tested based on algorithms, images and workstation conditions.
Enterprise projects also need to confirm whether the current environment is a development verification or formal commercial environment. The developer kit is suitable for research and development and adaptation verification. For formal delivery, the commercial software license, mass production hardware form, runtime acquisition channels and subsequent upgrade support should be confirmed.
Step 2: Connect the camera and fix the imaging conditions
The input quality of the vision algorithm is determined by the camera, lens, light source, mounting position and workpiece status. Just because the camera can be recognized by the system does not mean that the imaging has met the detection requirements.
Adaptation work usually includes:
- Identify USB/UVC, MIPI or network camera devices;
- Confirm driver, device permissions and service running user;
- Verify resolution, frame rate, frame fetch timeout and disconnection recovery;
- Fixed field of view, working distance, focal length and workpiece position;
- Fixed exposure, gain, white balance and light source conditions;
- Confirm the pixel size of the smallest target or defect in the original image;
- Create UNKNOWN rules for blur, overexposure, underexposure and occlusion.
The current project of Winge Technology has connected the USB/UVC camera link on Atlas 200I DK A2 and formed a standard image collection and video detection entrance. The fully China-localized solution will further integrate Chinese cameras, lenses and light sources to complete the imaging verification of the complete system.
Step 3: Choose an algorithmic route suitable for the workstation
Fixed-station visual inspection does not always require deep learning models. Algorithm selection should be determined based on defect definition, sample number, pose variation, and imaging conditions.
reference image and traditional vision
It is suitable for workstations with stable position, relatively fixed normal appearance and few samples. Common processes include global registration, ROI cropping, brightness normalization, local difference calculation and geometric rules. This type of algorithm is convenient for outputting difference locations and evidence image. The current project has completed its operation verification on the Huawei Ascend Atlas device.
Deep learning algorithm extensions
It is suitable for scenes with many categories, large pose changes, complex backgrounds or where semantic classification is required. It is necessary to prepare representative data, train object detection, classification or segmentation models, and complete input and output, pre- and post-processing, result consistency and target platform deployment verification.
Without real data and clear accuracy goals, simply completing a model deployment cannot demonstrate field detection capabilities.
Step 4: Establish reference image, ROI and detection rules
Assembly inspection requires translating quality requirements into executable rules. For example:
- where the part must exist;
- What is the correct quantity;
- Which categories fall under mispackage;
- How much variation is allowed in position and orientation;
- What to do when the image quality does not meet the conditions;
- How to summarize multiple rules into PASS, FAIL or UNKNOWN.
Winge Technology The current software supports collecting multiple OK reference image, delineating ROI, configuring missing/wrong-part assembly rules and saving detection evidence. Multiple reference image are used to estimate normal fluctuations, but 3 reference image only reach the image number threshold set by the current code. The formal workstation should still cover batches, shifts, temperatures and allowed imaging changes.
Step 5: Complete the operation interface and result evidence
The engineering site not only requires algorithm output, but also requires operators to be able to configure and review it. Adaptation interface usually includes:
- Camera and image source selection;
- OK reference image collection;
- ROI and rule configuration;
- Single, continuous or triggered detection;
- PASS, FAIL, UNKNOWN status;
- Original drawings, evidence drawings, rules and time records;
- Chinese, English or other language interface;
- Operator, Engineer and Administrator permissions.
The current project has formed a Chinese and English Web configuration platform, and verified the main pages, cameras and result links on the Atlas actual machine. The official commercial version also needs to add login, permissions, TLS, auditing and recipe approval according to customer requirements.
Step 6: Access I/O, PLC or business system
The visual inspection results need to correspond one-to-one with the actual workpiece. Field adaptation should specify:
- who sends the workpiece into place and detects the trigger signal;
- When does the camera take the frame;
- How to deal with timeouts, repeated triggers and missed triggers;
- How to output PASS, FAIL and UNKNOWN;
- How to intercept or transfer unqualified products to manual review;
- How MES, barcode or serial number correlates results;
- How to recover after a service restart or network outage.
The fully China-localized solution will use Chinese I/O or control equipment for connection. Before the interface joint debugging is completed, it can only mean that the software has interface design capabilities, but it cannot mean that the on-site closed loop has been accepted.
Step 7: Establish deployment, upgrade and rollback capabilities
The development environment can directly modify the code, while the production environment requires traceable versions. It is recommended to keep at least:
- Software version number and release date;
- Release package and SHA-256;
- Recipe and configuration backup;
- Current version and previous version pointers;
- Service status and key file hashes;
- Upgraded cameras, interfaces and smoke detectors;
- Explicit rollback method.
Winge Technology The current Atlas project has established an immutable release directory, shared data, configuration backup and rollback points. This capability solves delivery and maintenance issues and does not mean that the accuracy of the algorithm has been accepted.
Step 8: Complete formal acceptance according to workstation
Formal acceptance should at least cover:
- Independent OK/NG sample set;
- Missing detection rate of key defects;
- OK false interception rate;
- UNKNOWN rate;
- Single piece tempo and resource usage;
- exposure, position, batch, and temperature changes;
- Continuous operation and recovery from camera disconnection;
- I/O, PLC or MES closed loop;
- Recipe switching, permissions and auditing.
The number of tests, thresholds and qualification standards are confirmed by both project parties based on product risk. Composite pictures, parameter adjustment samples or short-term smoke tests cannot be used as a substitute for production acceptance.
FAQ
Does starting the software on Atlas mean that it has completed the adaptation?
Not equal to. Also verify cameras, algorithms, rules, results, interfaces, deployment and target station acceptance.
Is the detection software actually running on the Atlas device?
Already running. The current project has completed software startup, USB/UVC camera access, standard images, ROI rules, Chinese and English interfaces, detection results and version rollback verification on Atlas 200I DK A2.
After the adaptation is completed, can it be copied directly to other workstations?
It cannot be directly assumed. Different cameras, lenses, light sources, workpieces and beats will change the input conditions, and each workstation still needs to be reconfigured and accepted.
Can the Atlas developer kit be directly used as a mass production baseline?
You should first check the software license, commercial operating environment, mass production hardware form and support channels before determining the formal delivery plan.
Sources and explanations
- Atlas 200I DK A2Basic specifications:https://www.hiascend.com/document/detail/zh/Atlas200IDKA2DeveloperKit/23.0.RC2/pd/pd_0004.html
- Atlas 200I DK A2 software usage statement: https://www.hiascend.com/document/detail/zh/Atlas200IDKA2DeveloperKit/23.0.RC2/qs/qs_0018.html
Related Atlas vision inspection resources
- Fully China-localized assembly vision inspection: Huawei Ascend Atlas project case study
- Winge Technology has completed on-device adaptation of assembly vision inspection on Atlas 200I DK A2
- From Atlas developer-kit validation to a China-localized production vision system
- How does a Chinese edge platform support assembly completeness and missing-part inspection?
- How to deploy and deliver Huawei Ascend Atlas assembly vision inspection locally
- Huawei Ascend Atlas Vision Inspection
Online
Phone
WeChat
Top