Training Algorithms

This is applied when detection and localisation of instances of visual objects (i.e. defects) of certain classes is required. Image datasets of parts with and without defects are used to train this model.

A class can be created to label each defect/object in an image using a bounding box label to identify and assign it to a target class. The images without defects do not need to be labelled with a class but should be included in the image dataset.

Once trained, the object detection model gives a result of a bounding box around the recognised defect or object in an image. Object detection can be used to produce a test result that will give a count or a pass/fail output.

Manufacturing companies can use object detection to spot defects or missing components in the production line.

This type of algorithm is used when localisation of defects is not necessary. Classification recognizes, understands, and groups an object into specified classes.

Generating large amounts of labelled data is very quick and easy, as it only requires assigning a label to an image.

The deep learning neural network extracts features from images that can then be used to classify the image into a prescribed class.

A classification model can be used to classify images of a parts into categories, for example for part A or part A-B variant.

This can be deployed to train a model to detect and localise irregular shaped objects. It can define the exact outline of an image object using a polygon outline, as opposed to the bounding box labelling used in object detection.

The instance segmentation algorithm works in a manner where it will ignore the background of an image. The goal of this training model is to get a view of objects of the same class divided into different instances. Instance segmentation works by using a combination of classification and object detection.

A class can be created to label each object in an image using a polygon outline of each class, as well as each instance of particular class. The images without defects do not need to be labelled with a class but should be included in the image dataset.

The advantage of this model is that it can detect and measure pixel-level objects. The resulting output of segmentation should effectively be a mask that outlines the shape of an object. Instance segmentation can be used to produce a test result that will give a count of each class instance or a pass/fail output.

This model is used in the electronics industry when verifying the presence or absence of conformal coating in certain areas on PCBs.

Anomaly detection can be utilised for the detection of anomalies or outliers that have not been included in the training data set i.e., abnormal patterns that vary in shape, colour and size that are outside of the expected range of variation of “good samples”.

Two sets of image classes such as “good” or “bad” should be assigned without the need to label specific anomalies.

The image level detection focuses on the question of whether the whole image is good (normal) or bad (normal).

An anomaly score is generated to indicate the severity of an anomaly. The lower values indicate the absence of anomalies and higher values indicate the presence of high-intensity anomalies.

The anomaly detection algorithm learns what is typical or expected patterns from the normal data images. The abnormal image dataset is used to determine a threshold where an anomaly score above that threshold indicates an anomalous sample.

Inferencing or testing will show passes for normal samples and will mark the anomalous regions of an abnormal sample by means of a heat map.