AI: anomalies in component photographs

Serial details can be the same in the drawing and significantly differ in the photos. The trace of the treatment, glare or acceptable shade of the coating is easily confused with chipping and extraneous inclusion. When there are a lot of pictures, the buyer needs a clear way to highlight suspicious products for a specialist, preserving the original frame and the reason for sending for inspection. A local AI server for visual control makes sense to select exactly for such a process, starting with one group of parts and a verifiable result.

Request a quote
AI Server Manager 256
Illustration of a server platform prepared with the help of AI from the original photo. The exact scope of supply is determined by the specification.

You can request RadarStation. AI Server Manager 256 It's called Superintellect, which is the name of the catalog quotation, and 256 is the designation of the model, which is not confirmed memory. Configuration of equipment, software and work agreed on the task. To begin with, it is more useful to send several characteristic series of images and a description of control than to choose a server by one name.

Start with a product that really repeats itself.

Suppose a company takes pictures of treated enclosures before packaging. The controller is interested in visible surface damage, unusual inclusions and build differences. For the first stage, a specific family is selected, the necessary sides of the case and the conditions under which the product is considered fit. The employee of the quality department confirms the samples of the norm and separately notes cases requiring additional inspection.

The working folder should contain the permissible variety: batches of material, normal traces of processing, options for allowed marking. If you collect only perfectly lit advertising shots, the usual production batch will be unexpected. Together with the photo, the details ID, date, batch and specialist decision are stored. These connections allow us to understand later what conditions gave unnecessary alarms, without turning each dispute into a manual search through the general archive.

Normal sampling and difference model

One of the candidates for the pilot is PatchCore in Anomalib. The official description explains the approach through the feature bank of fragments of normal training images: fragments of the new image are compared with it, the result is an assessment of the anomaly and a map of differences. This is a useful starting point when there are more confirmed normal examples than photos of all possible defects. The card helps to direct the attention of the inspector, but does not establish the reason for the marriage. PatchCore documentation.

Before checking the model, it is necessary to establish the reception of the frame: the same kind of part, the readable work area, the controlled light and the separate state "reshoot". A blurry image or another person's modification must not automatically become a manufacturing defect. With several variants of the product, it is decided in advance which are checked together and which one needs its own sample. This design solution is tested on the examples of the buyer.

Threshold is selected together with the load on the controller

A system that sends almost everything for manual inspection may look sensitive and do little to help the job. The reverse extreme is a convenient short queue with unnoticed important cases. Therefore, acceptance is discussed in understandable terms: how many normal products caused alarm, what known defects were missed, how long it took to consider the disputed batch.

In the Anomalib PostProcessor manual, the threshold is associated with the test sample; normal and anomalous examples with known answers are preferred. The project fixes the selected threshold before the final test and saves the model version. Setting sensitivity on some frames and the final assessment on others allow you to see the benefits without fitting to already familiar answers. Handbook for results management.

Situation in the pictureWhat to check in the pilotWorking action
Permissible trace of processingIs it presented in a confirmed norm?Save the decision of the specialist and consider the option
Known damage.Does the desired area fall into the map and queueSend to the scheduled inspection
Glitter, blurred frameDoes the image pass fitness control?Request a reshoot.
New modification of the bodyDoes it belong to the current group?Check separately before project expansion

Next batch check

The same case, photographed several times, it is better not to divide between training and the final test: such a test is too similar to recognizing a familiar subject. It is more practical to postpone individual products or the next batch. In the control set, ordinary, borderline and confirmed defective cases are retained, and the responses of the specialist are opened after processing.

For a conventional batch of a thousand normal products, fifty alarms mean fifty additional examinations. This is an example of the load, not an indicator of the proposed solution. It is discussed together with the result on known defects: the average "accuracy" on mostly usable details can hide the problem. In the final protocol, real examples of errors, the scope of the test and the shooting conditions are useful. Internal serviceability and dimensions of the product are evaluated by the provided means of control, and not by the calm color of the card.

Server for photos and working window

For selection, you need the resolution of images, the number of types per product, the volume of the daily batch, the size of the archive to be stored and the permissible processing time. On the agreed set, the selected software, model and hardware are checked; then the CPU, RAM, drives and the need for a GPU are determined. From the image of the body or the number 256, these characteristics do not follow. Data preparation, configuration, controller’s workplace and maintenance are listed under separate quotation positions.

The first result of the project is conveniently limited to the list of products for inspection, initial images, difference maps and employee decisions. Automatic line control or final rejection requires separate approval. The price and availability of the platform – on request. In the application, specify the delivery city, for example, Moscow, St. Petersburg or Almaty, the conditions of accommodation and the required period: we agree on the server and the composition of work for your enterprise.

From photography to product solution
  1. Identified product and suitable picture
  2. Selected group of norms and model version
  3. Assessment and difference map
  4. Suspicious cases for examination by a specialist

A glare, blurred frame or an incorrect look Repeat shooting, separate mark about unusable entrance

Permissible change confirmed Save the solution; change the sample to check a separate version

Project organization of preliminary control. The anomaly map does not measure the dimensions and does not establish the internal health or cause of the marriage.

Questions before buying

The system will determine the cause of the marriage and block the shipment?

In the described project, the result is suspicious images and a difference map for a specialist. The reasons, final decision and management of shipments relate to separate agreed procedures; an anomaly may turn out to be a permissible change.

Can I start if there are few photos of known defects?

Yes, we can discuss a pilot with a confirmed normal sample. But for the assessment of passes, known defective and borderline examples are needed. Testing only on normal products does not show the ability to detect defects.

AI Server Manager 256 already has the right software and 256 GB of memory installed?

There is no such confirmation in the current quotation. 256 is the model designation. Components, memory, software, data, work and test criteria are recorded in the specification; price and availability are confirmed upon request.

The main product in our catalog

AI Server Manager 256

AI Server Manager 256

Price on request. Configuration, availability, software and work are agreed in the quotation. The designation of the model does not determine the memory capacity or composition of the processors and accelerators.

Request a quote

Documentation: Anomalib: PatchCore, Official Stable Documentation; Anomalib: Post-processing in Anomalib. These provisions relate to these editions; the applicability to the proposed configuration is specified separately.

All technical guide RadarStation