Deep Learning Automatic Inspection System Neural Networks Aoi Vision

Deep Learning Automatic Inspection System Neural Networks Aoi Vision Machine

Deep learning image processing is to solve image processing problems by building a "brain-like neural network" and drawing on the method of

human brain processing data. Therefore, any scene involving deep learning requires two necessary conditions: one is the support of big data— Various representative pictures; the second is a powerful computing platform - generally using a GPU computing platform. In actual industrial application scenarios, these two conditions are extremely difficult to achieve, and the implementation cost is extremely high. KEYE TECH AI is a deep learning function launched on the basis of the Kvis general development platform. It combines KEYE TECH's own powerful traditional image processing algorithms to solve most of the above two problems.

Equipmet configurations

The inspection machine adopts high-pixel industrial cameras and high-performance stroboscopic light sources to carry out omni-directional visual inspections for product appearance defects. The equipment can realize 7*24 hours of all-weather operation, and it can be online automatically. Eliminate substandard products.

Application range

Advantages and Features
(1) Neural network inference is completely implemented based on CPU: The neural network model inference trained by KEYE TECH AI is completely implemented based on CPU, and the neural network model training supports both CPU and GPU. Users can flexibly choose a computing power platform according to the complexity of the model required by the project, thereby saving unnecessary hardware expenditures on the computing power platform;
(2) Innovative "deep learning defect filtering" algorithm: use traditional algorithms to find all possible defect areas, and only use deep learning to solve the OK and NG discrimination of defect areas, thus solving the problem of training image data acquisition;
(3) Support the retraining of neural network models: According to the complexity of the application of industrial scenarios, the continuous training of models is supported, so that industry-specific neural network models can be formed.

Precautions

1. The training data set must strictly focus on "quality" and "quantity"

•  Quality: The data set used for training must be representative, contain various possible situations, and the distribution of the number of pictures corresponding to each situation should be uniform.
• Quantity: The number of data sets used for training must be relatively large. According to the complexity of the model, the number of high-quality pictures must be large.

2. Traditional algorithms should be combined with deep learning. Traditional algorithms can do it, don't rely on deep learning

• Traditional algorithms: positioning, size, etc.
• Deep learning: refer to the “problems that are difficult to solve by traditional algorithms” introduced earlier.

3. Accurately locate specific problems, from "partial" to "whole". The ultimate goal of a machine vision project is to achieve 100% accurate detection, but there will be various issues during project development that affect this result. When an abnormal result occurs, it is necessary to accurately locate the specific link to test and find.

• Review of the training data set: After the data is classified, it needs to be checked again after labeling to see if there is any labeling error or classification error.
• Neural network model reasoning: After model training, be sure to test a large number of pictures before importing engineering files for online reasoning.

After-sale service

The company has a complete technical service team and rapid response mechanism, and has dedicated service specialists for each customer, who can receive technical consultation and fault reports from customers at any time. And to ensure rapid response to customer emergencies, to ensure that customers receive satisfactory service.During the epidemic or due to special reasons, when after-sales engineers are unable to reach the site, the service center can remotely adjust customer equipment for troubleshooting and technical consultation.

After the equipment arrives at the customer site, the after-sales engineer arrives in time to carry out equipment installation, commissioning, and operation training. The product quality of the whole machine is traceable, and the quality warranty period is 1 year from the date of acceptance. In the event of non-human faults during the warranty period, after-sales engineers will quickly arrive at the site or provide remote guidance for free maintenance.