Vision AI

Table of Contents


Vision AI is an automated anomaly detection platform based on computer vision and machine-learning algorithms. It enables computers and systems to derive meaningful information from visual inputs—and take actions or make recommendations based on that information. Built specifically for production facilities, it leverages deep-learning and AI vision technologies for product inspection, factory process automation, and quality control. Once the platform is fed with visual inputs, it identifies the region of interest on the product’s surface and represents the defects in the form of heat maps, error maps, and reconstructed images.

High Peak Software’s Vision AI platform carries out surface-level anomaly detection in products and generates production-ready solutions within minutes, catering to multiple industries such as automobile, apparel, footwear, healthcare, textile and more.

Why did High Peak see the need to create Vision AI

In manufacturing, current market volatility is leading to immense operational pressure. The industrial details are becoming finer, and manufacturers need to get their products to the market with less waste, faster, and on time. However, the conventional manual defect detection methods used in the manufacturing landscape often result in resource waste and issues with product delivery. Moreover, with no production line generating 100% indefectible and sound items, the traditional machinery inspection can cost manufacturers and retailers millions of dollars and thousands of hours.

As a result, manufacturers across the globe are struggling with inefficient production pipelines, resulting in increased demand, supply chain disruption, and interrupted retail channels. High Peak Software wanted to build a deep-learning driven defect detection automation platform to help manufacturing facilities:

  • Streamline the production lines’ output with zero defects and 100% quality products by accurately and efficiently inspecting every single instead of just random samples.
  • Train and generate defect detection models by leveraging cameras, data and algorithms in the production facilities. A system trained to inspect products or watch a production asset can analyze thousands of products or processes a minute, noticing imperceptible defects or issues in much less time compared to humans. 
  • Improve production units’ overall quality, accuracy or performance by overcoming variations across different human inspectors. Thus, downsizing manufacturing operational costs.
  • Reduce manual maintenance effort by deploying a fast, reliable, and highly accurate anomaly detection platform.

High Peak encountered and solved for several technical and business challenges during Vision AI’s design and development

1. Lack of quality data

Computer vision-based algorithms and systems need large and diversified datasets to be truly effective. Labelled and annotated datasets are critical for effective analysis of the visual information, training a model, and its deployment. There are lots of factors that can throw a system off. Shades, colouration, darkness and light, and odd shapes can potentially confuse the system, causing it to produce unexpected outputs. Though a well-trained system can deal with most of these issues, it might still fail to process and analyze something that it did not encounter before.

Our engineers at High Peak needed to leverage AI-based algorithms to build a system capable of learning context through visual data analysis autonomously and be able to recognize and differentiate between visual inputs in real-time.

2. Suboptimal computing resources

The boom in popularity of AI technologies (ML, DL, NLP, and computer vision) have largely commoditized access to best-in-class computer vision libraries and deep learning frameworks, distributed as open-source solutions. Coming up with an algorithmic solution to computer vision problems is no longer an issue. The modelling know-how is already available. It is computing power that often constrains model scaling and large-scale deployments. 

Since machine learning and deep learning algorithms are computationally demanding, one of the major challenges for our team was to design and develop an optimized memory architecture for faster memory access—stack up on CPUs and GPUs. Moreover, the platform should be able to run on local hardware and hybrid cloud platforms.

3. Managing stakeholders’ expectations on short project timelines and product use cases

When estimating the time-to-market for computer vision-based applications, some stakeholders overly focus on the model development timelines and forget to factor in the extra time needed for data collection, cleaning, and validation along with model training, testing, and deployment. Combined, these factors can significantly shift project timelines.  

Also, deploying industrial machine vision just to eliminate the human factor is not the right approach prevailing in the current market. These platforms can easily replace the quality control department of a production line only if there is an actual use case with quality visual inputs to be processed, analyzed, and deployed in order to get the best ROI.

How High Peak Software’s Vision AI is powering industries with production-ready solutions

Vision AI works the same way as human vision, except humans have a head start. Human sight has the advantage of lifetimes of context to train how to tell different objects apart, what kind of anomaly is present in an object, or whether there is something wrong in an image of that object. The platform process and analyze the visual inputs to train models to perform similar functions, but it can do it in much less time with data and algorithms rather than retinas, optic nerves and visual cortex. Because it is designed and trained to inspect thoughts of product images in no time, noticing unapparent anomalies, it quickly surpasses human capabilities. Consequently, driving business impact at scale by improving production units’ productivity.

Below, we’ve detailed the core features of the Vision AI platform and how it can be leveraged in production facilities across industries.

1. Real-time unsupervised anomaly detection with error-free visual inputs

Manufacturing plants involve specialized equipment for producing goods, which drastically reduces the chances of any regular error occurrence. Since production workflows are automated and machine-run, stakeholders don’t have visibility on what all can go wrong and when it can go wrong. Moreover, these systems don’t have enough qualitative and quantitive data to carry out a conventional/supervised training where they can guide the platform about an error or series of errors to look for in a product. If left unresolved, it can have major implications on a production facility’s overall output efficiency.

This is where Vision AI enables manufacturing plants with an unsupervised automated learning model. The platform trains the model with just 5-50 error-free product images. These images are then fed as training data to the model. By learning an object’s normal or optimal appearance, the model identifies unpredictable defects on the product’s surface in real-time. 

Auto Mode

2. Guided (supervised) anomaly detection by learning annotated bad areas in visual inputs

In the use cases where a product defect is known, the platform is fed with a few bad images and test good images for learning. These images are labelled for the platform to differentiate and understand the defect. 

Guided Mode

To guide the model into looking for specific defects, users can annotate or highlight each defect individually in multiple ways, namely by using a marker or drawing various shapes such as a rectangle, circle, or polygon around the region. Users can also annotate using free form drawing. The platform also provides users with an eraser for editing the annotation. Additionally, users can use the zoom in/zoom out feature for greater clarity.

Guided Mode with defect annotations

In addition, the original dimensions of images are retained, and each defect is provided with a different colour for easy classification. Based on the user’s selected image, a unique model is generated for every identified defect. This user validation data is then captured and used for generating model training data. By annotating the defects on the input images, the trained model looks for only these defects when provided with new images.

Model Validation

For both the use cases, identified defects (deviation from standard visual inputs) in the product are represented on heat maps, error maps, and reconstructed images. These models then can be evaluated by the users on model confidence score, training and validation accuracy, precision, F1 score, CPU and GPU memory usage, and the algorithm’s performance. On satisfactory results, these models are tuned with hyperparameters and can be deployed in production to access the product’s visual inputs in real-time.

Model Training Dashboard

Moreover, error-free and defective product images are stored in an image repository. These images are tagged and labelled accordingly for training purposes. Each time a new image is fed to the platform, a new model is generated for that particular image. These models get stored in the model repository for future references with their respective training results.

Model Evaluation

Vision AI offers an intuitive and smooth user experience with features like multi-user login and network-based multi-location login for easy access. Vision AI can be plugged with multiple third-party visual inspection softwares and is currently being leveraged to solve multiple problem types across industries for surface-level anomaly detection. 

Technologies used

  • Frontend – Angular / React Framework
  • Backend – Django/Flask (Python-based) Backend Server, Java-based Backend (NodeJs)
  • Deep Learning Framework – Tensor Flow 2.0 / Caffe
  • Communication Framework – Rest API based communication

Subscribe to our monthly newsletter