Augmentation Glossary
- Daniel Leal (Unlicensed)
Horizontal Flip & Vertical Flip
The argumentations Vertical Flip and Horizontal Flip are recommended when the target images are not standardized/homogeneous part (FabricWood/Metallic Surface, bulk products, eg. visual inspection of pills on a conveyor, foreign body detection in a bulk pile, etc., using โHorizontal/Vertical Flipโ argumentation, may improve detection of defects.ย
It mirrors the images Vertically & Horizontally, comparing it with the original one in order to find differences, therefore improving accuracy.ย
However, this would not be recommended for instance, when inspecting defects on static visual inspection, where the inspection is focused on a specific target area of parts, such as PCB (Printed Circuit Boards), since when the part is flipped the detection area will be different than the target area.
Width shift (min 0, max 100)
A shift to the left or right will occur.
Height shift (min 0, max 100)
A shift up or down will occur.
Network type
The network type determines network accuracy and the amount of time necessary for the image to be evaluated. The faster the network, the less accurate it is. We ussually recommend to not modify this value, as 'balance' value has demonstrated to be efficient in most cases.
Color jittering (min 0, max 50)
Color jittering augmentation adjusts the color of RGB image,ย with a selected value. This argumentation may be suitable for cases where images have a wide range of colors or shades.
It has been effectively used in previous cases, where for example, the imageโs background was very dark and we were attempting to detect holes for a target part.
Number of training cycles (min 1)
The more, the better.
Averagely, 3-5 training cycles are enough for good results, however, sometimes up to 10 cycles may be required for optimal results.
If you increase the number of training cycles substantially, processing times will increase considerably
If the differences between OK and Defective images are not very significant, is recommended to increase the number of training cycles.
We recommend testing more training options and evaluating which model gives the best results.
Resistance to deviation (min 0, max 100)
Resistance to Deviation can be described as a fine tune tool, when its value is increased it will ignore detection of small defects. This argumentation should be implemented, when thereโs presence of a lot of false positives. In sumarry, it can be used to ignore noise or other deviation in the image.
Attention: Please note that other settings may be the actual cause of false positives. Therefore, only attempt to change this argumentation value when you made sure the correct best practices were done overall for your project.)
Brightness resistance (min 0, max 100)
The name itself is suggestive, itโs a value to alter the resistance to brightness within the picture, in other words, itโs the resistance to changes in lighting.
Brightness Resistance values should be increased if you think the lighting is affecting the softwareโs ability to identify defects due the brightness/lightness on the image.
Shear (min 0, max 360)
Object can be skewed by a given factor in the direction of x or y axis.
Saturation Resistance (min 0, max 100)
Saturation is the strength or purity of the color and represents the amount of gray in proportion to the hue. A "saturated" color is pure and an "unsaturated" color has a large percentage of gray.