The client is a multinational product manufacturing company, which specializes in manufacturing interior design products for commercial and residential spaces. The company is specifically engaged in the business of manufacturing, selling and distribution of paints, coatings, products related to home décor, bath fittings and providing related services.
Project Objective – Defect Detection Using Image Processing with AutoML
Business Value – The client intended to identify product defects faster and deliver improved quality of products and services to their customers.
The client was looking to automate identifying defects on the various surfaces their paint products are used on, such as painted areas, paint solution found in containers, paint samples on the hand, etc. They were particularly interested in defect detection using image processing and AutoML solutions in the cloud, in order to have no dependency on the domain knowledge and the skill set required for it.
GCP’s AutoML for image multi-label classification can provide a low-code solution that allows users to search through a zoo of ML algorithms, and leverage a single model for multiple defect types that get detected with respective confidence thresholds.
This AutoML service is part of the GCP’s Vertex AI platform that is an end-to-end ML platform, which allows for deployment of the AutoML models on managed Google Kubernetes Engine clusters.
The solution helped the client to identify defects based on categories and criterias. This included whether the defect is above/below/on the surface, defect color/texture/physical. The categories for defects included algae, blisters & bubbles, sand-like particles, efflorescence, patchiness, peeling, poor adhesion, rough finish, shade fading, and more.
- The images shared by the client were categorized into specific defect categories, to be fed to the model for training.
- The categorization was performed on a specific criteria. Based on the categorization of the defect, folder name spaces are used for defect specific images
- Those images which have more than a single defect were used in corresponding defect folders of the input training data.
- The image dataset .zip file was uploaded to the respective cloud storage buckets.
- The zipped files stored in Google Cloud Storage buckets were imported into the Vision AI / Vertex AI multi-label image classification models for training.
- In Vertex AI UI Console, the respective labeling/annotation for the defect types was performed.
- Vertex AI/AutoML enabled the above feature engineering by being able to track the lineage between Datasets, their training iterations, Models generated and the eventual deployments.
Vertex AI also supports deployments to Endpoints that run on managed GKE clusters, without the need to maintain them.