AIST, one of Japan’s largest research institutes, invented the feature extraction method combining
"High-order Local Auto-Correlation (HLAC)" and multivariate data analysis.
Adacotech has implemented this method into its original anomaly detection system and is applying it to various fields.
No need of Anomaly Data
Anomaly detection models can be trained from a small amount of normal data (unsupervised), compared with deep-learning methods.
No need of GPU
High speed processing is available on a normal PC due to simplified calculation.
Not "Black Box"
It linearly processes HLAC features, so the calculation process and result are explainable.
What is High-order Local Auto-Correlation (HLAC) feature extraction method
(AIST patent technology)?
HLAC is a feature extraction method used for the analysis and recognition of images, etc. It is versatile, high-speed, and has excellent recognition accuracy.
Since it can be calculated by product-sum operations, the shape and size of the object can be recognized instantly using a normal PC.
In addition, since it is position invariant (does not depend on the position of the object), segmentation is not required (segmentation free).
Furthermore, it has characteristics preferable for image recognition such as additivity
(if there are two objects in an image, the sum of the respective features result in the feature of the whole image).
What is judgement using HLAC?
After extracting the HLAC features of the data to be analyzed (image, video, sound, sensor data, etc.),
the range (subspace) of data that is considered "normal" is determined by principal component analysis,
and the distance of deviation from the subspace is determined as abnormal value.
Flow of HLAC judgement
Comparison of HLAC and Deep Learning
|Number of Samples
|More than 10,000
Normal case data only
|Normal case and defective case data
Transparent and traceable
|Time Required for Training
A few seconds to a few minutes.
|A few hours to a few days.
Normal PC ( -$1,000)
|Set of highperformance
GPUs ( +$100K)
Nearly false negative 0%
|Very difficult to achieve
false negative 0%