Deep Learning AI Defect Detection For Industrial Product Lines

Industrial graded defect detection with precise location marked on products; runing on Jetson NX and other edge devices. Automatically tell the defects location on the product for industrial product line quality control.

Background

Most of the algorithms in the AI field of defect detection require a large amount of good and bad sample dataset. Those data (especially the bad sample data) highly impact the performance due to unpredictable defects often occurring in product lines due to machine failures, human mistakes or more. If any of the defect cases are not included in the given bad samples during the training time, the AI model may not predict the defects correctly.

 

In order to train the AI model with the new defect cases, it will cost a tremendous amount of time to collect the data and find a way to repeat the process which causes the defects, also retrain the model. This process not only costs the time of the factory and also requires unnecessary time to improve the model over time, and this process will happen again if new unpredictable defects occur. Therefore we introduce our defect detection technique that differs from other traditional machine learning methods, anomaly detection, which achieves the Image-level anomaly detection accuracy (ROCAUC) of 95% or above in the MVTec Anomaly Detection Dataset. 

Project scope

Our technique only requires good sample data which is easily collected by the factory in a controllable environment. We take the given good sample data and extract features in different perspectives using machine learning feature extraction techniques. Although they are good products, they still have slight differences between them; the feature extraction technique is able to train the model to learn the acceptable differences between the good products and those cases should not be counted as defects.

 

During the deployment, we take the testing images to compare with the features learned by the model and extract the differences. The feature differences are represented as a map which is able to tell us the defect locations precisely on the defect products. 

 

We used only 40 good samples to extract our features for demo purposes. This demo, trained with only good samples of a chip, tries to identify the defect products and the defect locations of the products. With 40 training images each with size of 600x600 pixels, we are able to find all the defects precisely on the given test-set.

AI deep learning model on Edge device

With Nvidia Jetson-NX Edge device we achieved ~15FPs 

Results:
The following is an example of a training image (left), test image (mid), and output (right) of our defect detection technique:

defect_good.jpg
defect_bad.jpg
defect_marked.jpg

As shown in the above figure, the model is able to find multiple defects on the chip and localizes them correctly.

Environment and conditions required

As the model is for extracting the defects of the product, the operating environment should be stable and controllable.

- the light source should be fixed and consistent with single color, intensity and direction

- the background should be single colored and as clean as possible

- the product to be detected should be in the same position and location

- small variations of the product (e.g. the serial number on the product) can be trained to be ignored by the algorithm in case the variations are included in the training set samples.