SaigeVision

Image Inspection Toolkit

Automate your vision inspection tasks with the powerful, reliable, and easy-to-use SaigeVision® platform. Employing the state-of-the-art in machine learning algorithms, with advanced tools for data generation, augmentation, and labelling, SaigeVision® offers unmatched accuracy and speed for the most challenging inspection problems. Even with only limited training data, SaigeVision® achieves maximum performance with minimum operational overhead.

Features

Fast

Using a combination of optimal transfer learning to construct sparse networks, advanced model compression and pruning techniques, and code and algorithm optimization, SaigeVision® achieves best-in-class inspection speeds. Training times are also greatly reduced via focused attention learning.

Flexible

SaigeVision® is designed to be intuitive, modular, scalable, and easily integrated and deployed. Various labelling and support tools considerably ease the user’s burden in preparing and training the network. No expertise in rule-based vision systems and programming is required.

Accurate

SaigeVision® achieves best-in-class accuracy through a combination of our proprietary image generation technology, auto-labelling, and other advanced machine learning techniques. Both false negatives and false positives are significantly reduced.

Main Functions

  • Classification

    Quickly determine whether a sample is normal or defective
  • Detection

    Determine both the type and location of defects (using bounding boxes)
  • Segmentation

    Determine both the type and location of defects (pixel-level contours)
  • Image Generation

    Generate realistic synthetic images to be used for network training

Specifications

Level

Minimum

Recommended

Operating System

Windows 7 64-bit / Windows 10 64-bit

CPU

Intel Core i5

Intel Core i7

Environment

Visual Studio 2013

Visual Studio 2013

RAM

16GB

32GB

GPU

NVIDIA GeForce RTX 3060, 3070

NVIDIA GeForce RTX 3080, 3090

HDD

 

SSD

Key Features

1. Optimal Transfer Learning and Directed Attention Learning

  • By designing a transfer learning algorithm that optimally transfers the weights of an existing network, training times are considerably reduced while producing a sparse, easily pruned network
  • Our directed attention learning algorithm focuses learning on the more important defective regions

2. Image Data Generation Technology

  • The Problem of Insufficient Data
    Deep learning networks perform only as well as the data with which they are trained. The more the data, and the richer the data, the better the performance. Contrary to expectation, most users have very limited data, especially of defective samples.
  • Image Data Generation Technology
    Generating synthetic data is trickier than it seems: simple manipulations of existing image data, e.g., flipping, rotating, blurring and other distortions, are not helpful. Generating realistic defective images requires more sophisticated techniques.
  • GAN-Based Image Data Generation
    SaigeVision® makes use of advanced GAN methods developed in-house to generate realistic synthetic data customized for industrial image inspection applications.

Benefits

Benefit 1.

Improved Inspection Accuracy

With the addition of high quality balanced training data, inspection performance can be significantly improved. An EV battery production line reported a 77% reduction in false negatives using the GAN synthetic image generation technology of SaigeVision®.
*False Positive: A defect incorrectly identified as being non-defective
Benefit 2.

Rapid Deployment

Data for new products and new lines takes time to accumulate, but data from related products and lines is usually available. Synthetic images based on related product data can be generated and used as training data for the new product/line, allowing for rapid deployment.

Using Image Data Generation for Inspection of a New Product or Line

  • Use related product data to train a GAN to generate synthetic images

  • Use the trained GAN to generate synthetic data for a new product

  • Train the inspection network for a new product using the generated synthetic data images

  • Deploy for inspection of a new product

  • Generate Realistic and Accurate Synthetic Data
    SaigeVision® takes advantage of generative adversarial networks (GAN) to generate synthetic images of defects that are virtually indistinguishable from actual defects, even by human experts. A wide range of defect types and sizes can be generated at arbitrary locations while preserving background homogeneity.
  • Reliable and Robust GAN Training
    GAN methods are notoriously difficult to use in practice; training a GAN can be time-consuming, and highly sensitive to mode collapse and other instabilities. SaigeVision® uses in-house algorithms optimized for industrial inspection that overcomes these and other challenges, delivering state-of-the-art performance for real scenarios.
  • Rapid Deployment for New Products/Lines
    Deep learning solutions are only as good as the quality and quantity of the data with which they are trained. Using our GAN-generated image data, inspection can be performed even for new products and lines for which data is not yet available.

3. Auto-Labeling

  • For most industrial inspection tasks, the data with which the deep learning network is trained is labelled by human experts. This can be time-consuming, and humans are prone to fatigue and misjudgment. Since most of the training data is from non-defective samples, auto-labelling can be used to identify those images that are non-defective, separating these from the much smaller number of images that are likely to contain defects. Human experts then only need to examine and label this much smaller set of images.