There are several factors which can affect the quantity of samples required for identification of defects on a specific product. Mainly, we higlight the following factors: quality of images (camera resolution, shadow, lightning), type of material, how many classes/types of defects are required to identify and others.
However, as a general advice, at least 20 GOOD and 20 BAD images, that should be enough to run a first successful attempt with our software.
Ideally, images should be provided on 2 separate folders, labelled for example, GOOD and BAD, or alternatively, files with prefix/suffix GOOD/BAD.
Please note below best practices when requesting a feasibility study request:
Provide enough quantity of pictures which would cover the majority of defects for that product
Provide defect types/class names for accurate reporting
Provide acceptable defect threshold (scratch size, hole diameter.. for OK/NOK scenarios)
Please note if the initial results are not satisfactory on the first attempt, more samples may be required depending on the product, therefore, we recommend to provide as many samples as possible.
If you have further queriesIn order to process the feasibility study on your provided images, PEKAT VISION requires several rules and information.
Image Quality Requirements
➔ Pictures taken by industrial camera with a minimum of 500 x 500 pixels depending on the complexity of a project. (e.g. not smartphone camera image)
➔ Objects captured in constant conditions - constant light, constant camera position and constant camera settings.
➔ Defects must be clearly visible.
Dataset Requirements
➔ Minimum 15 pictures of a defective object with well visible defect (“defective picture”) and at least 15 pictures of a non-defective object (“non-defective picture”) depending on the complexity of a project.
➔ Pictures divided intofolders OK and NOK sorted by defective and non-defective pictures.
➔ Minimum of 5 images of the same type of defect depending on the complexity of a project.
◆ For example, considering a color stain and unwelcome object on a textile - the dataset contains 5 different pictures of a color stain defect and 5 different pictures of an unwelcome object.
➔ Documentation that describes the defects, so it can be easily determined what is the real defect in the picture (and what is not). The best option is to provide some image samples with manually marked (circled, highlighted) defects.
➔ Cycle time – how much time we must evaluate the defect on a product or how many products pass through the line per second.
Please, keep in mind that the feasibility study report might take more than one week to process. If you have questions, please contact support@pekatvision.com