1 Upload training data to Amazon Lookout For Vision or

Serverless Architecture for Product Defect Detection Using Computer Vision import training data from Amazon Simple Storage
Service (Amazon S3) and train a model.
Detect product defects, get real-time notifications, and visualize insights using AWS artificial intelligence and machine
Admin users signup and login to a management front
learning, and serverless services 2 end website.
Architecture for camera-based in-line or end-of-line quality inspection. Supports automated or one-time anomaly detection using image
Start or stop the model or do one-time defect
classification in the cloud; real-time monitoring and notifications; and analytics and insights from the classification results. 3
detection by uploading an image.
Camera or a client application invokes an Amazon API
4
Factory AWS Cloud Gateway endpoint to get a signed URL from Amazon
S3. The request is authorized by an AWS Lambda
Amazon API Gateway AWS Lambda Image ingestion and storage function and a signed URL is returned.
4 Lambda function
Using the signed URL, an image, along with its
AWS Lambda Amazon S3 5
/designedly associated metadata, is uploaded to Amazon S3.
CameraId Custom authorizer Get signed URL Image uploaded event
AssemblyLineId Image upload to the S3 bucket triggers an event
ImageId 6 notification to initiate an AWS Step Functions
Camera or Client 6 workflow.
Upload images for inference Start workflow
5
Fetch image from the S3 bucket and present to
7
Management Front End Amazon Lookout For Vision Defect Detection Workflow Amazon Lookout For Vision for anomaly detection
2 using DetectAnomalies API.
/startModel AWS Step
Functions workflow Store inference result and image metadata in Amazon
Amazon CloudFront Amazon S3 /stopModel 8
/detectAnomalies DynamoDB.
3 Running 7 Publish a notification to an Amazon Simple
Model 9
Notification Service (Amazon SNS) topic to send an
Data scientist or admin Lambda function AWS Lambda
Amazon API Gateway Model
DetectAnomalies email alert to subscribed operators and plant managers
Train model
in case a defect or a low-confidence result is detected.
1 Persistence
Analytics Inference results are fetched from DynamoDB
DynamoDB 10
Streams, transformed and enriched, sent to Amazon
Streams
12 11 10 8 Kinesis Data Firehose for batching, and saved in
another S3 bucket.
Amazon QuickSight Amazon S3 Amazon Kinesis Lambda function Amazon DynamoDB AWS Lambda
Business user Data Firehose for transformation PutItemInDynamoDb Inference results datasets are imported into Amazon
11
Quicksight. This process can be scheduled based on
Monitoring and Alerting Amazon Simple requirements.
Notification Service (SNS)
Defects threshold exceeded alert Create dashboards and analyses for business users, and
12
gain insights from the inference results.
14 13 9
Amazon CloudWatch provides a single pane of glass to
13
Line operator or operators and plant managers for workload and defect
CloudWatch CloudWatch CloudWatch Amazon CloudWatch SNS topic AWS Lambda
plant manager Alarm(s) dashboard Logs PublishMessageToSNS detection monitoring using logs, alarms, and
Defect detected email notification dashboards.
Alarm notifications from Amazon CloudWatch are sent
14
to operators and plant managers by Amazon SNS
Reviewed for technical accuracy November 22, 2021
© 2021, Amazon Web Services, Inc. or its affiliates. All rights reserved. AWS Reference Architecture whenever defects exceed a pre-defined threshold.