Is it possible to detect this type of anomalies with only one model ?

[archocron]

Hello everybody, and thanks for this awesome library, Im working on a project, I would like to know from the experts if this library is suitable to solve a problem. The idea is to detect noise (like last image) inside different shapes in white (rectangles, squares, circles, and composed shapes) . Shapes can be in any size and any position.

First two images are OK, last one is Anomalous (note the noise in the smaller square).

Do you think is posible to use this library for this task training only one model ? I have trained one single model with all images, and the predictions are not accurate using Patchcore and around 100 goods images and 40 anomalous images.

Maybe any tips to get a better performance ?

Im using anomalib 2.0.0b3

Many thanks !

[abc-125]

Hello, this particular task looks like it is made for classic computer vision methods (detect white rectangles and see if there is any noise). In my opinion, it is quite different from tasks that can be solved with anomalib.
You can probably try to increase the number of good samples to 1000 or more to cover more cases, but I am not sure if this will be sufficient.