From Unattainable Dream to Industry Norm: AI Forges Its Path in Visual Inspection

Is AI used in your equipment?

Do you have your own deep learning algorithms?

Once, using AI was an unattainable dream, and applying AI to visual inspection in the industrial field was even more of a fantasy. Today, however, AI-powered visual inspection has become an industry standard.

AI Algorithms: Four Core Modules to Build an AI System

In 2017, when industrial visual inspection was still relying on traditional algorithms, Keye Tech took the lead in the plastic packaging industry to attempt using AI algorithms to solve defects that could not be detected by traditional algorithms.

Example:

In the visual inspection of brown graduated bottles, "black spots" have "unclear boundaries" and there are too many surrounding interference factors, leading to reduced detection accuracy and a high probability of missed detections with traditional algorithms.

Based on "data-driven" principles, AI algorithms effectively solve the problem of visual inspection accuracy in "interfering environments". By performing accurate data annotation (extracting features of the products to be inspected) on a sufficient amount of collected image data and providing a learning objective, computers can independently find patterns and summarize models from the data to generate their own "rules" (i.e., models).

To better implement AI algorithms in industrial quality inspection scenarios, Keye Tech has improved its AI algorithm system by establishing four core modules: image classification, object detection, defect segmentation, and geometric positioning. This has increased the original visual inspection accuracy by 30%.

Image Classification:

It can quickly determine the category of the entire image by learning features such as texture, shape, contrast, and grayscale values in the image. In NPU mode, the inference time per image is still as low as 1ms.

Object Detection:

It locates the target in the image against complex backgrounds, eliminates interference factors, and usually marks the target with a rectangular box.

Defect Segmentation:

It can classify pixels in the image to accurately identify the location and type of defects. It outputs key quantitative parameters such as the area of the defect on the image and the length and width of the minimum bounding rectangle.

Geometric Positioning:

It is often used in scenarios with complex image backgrounds such as diverse targets and postures. Through AI learning, it judges the shape of multi-endpoint targets and locates the target area. The error can even be reduced to within 3 pixels.

Example: Compared with traditional algorithms, AI algorithms have effectively improved detection accuracy by 30% in the inspection of lid R-angle defects, bottle body scratches, bottle mouth defects, and thread defects.

AI Empowerment: Improve Quality and Efficiency

In the process of AI algorithm deployment, to optimize the effect of the AI model, it is necessary to pay attention to model indicators and evaluate model performance by observing the data reflected in the confusion matrix. On this basis, targeted improvement measures can be taken, such as selecting and adjusting parameters that have a significant impact on model performance, adding new image data to the training set, and checking image annotation quality. Through multiple rounds of model training and optimization, the detection performance and effect of the model can be gradually improved.

It is worth noting that AI algorithms require high-quality image data. There are five key factors in collecting high-quality image data: accurate data annotation, balanced distribution of defect data, sufficient data features, adequate data volume, and real-time data updates.

To reduce data volume and training time, improve detection accuracy, and ensure ideal results during model deployment, Keye Tech has successively established the KVIS cloud training platform and built a cloud data center, while continuously optimizing algorithms.

Intelligent Annotation – Complete Annotation in 1 Second 

To achieve the interactive experience of "what you see is what you label", Keye Tech uses an intelligent annotation tool based on a visual large model. For defects with clear boundaries, annotation can be completed in 1 second, accelerating model construction. It is especially suitable for scenarios with clear boundaries and high contrast.

Positive Sample Training – Modeling Time Reduced by 50%

Collecting defect samples in the industrial field is time-consuming, inefficient, and may involve various unpredictable abnormalities. Keye Tech has launched a positive sample training mode, which can identify defective products only by using the features of good product images. It has an excellent foolproof effect and improves the detection accuracy of unknown defects.

Defect Generation – Modeling Efficiency Improved by 50%

The application of AI algorithms in visual inspection requires a sufficient training set. However, some defects are difficult to collect in actual production. Based on the current status of industrial scenario applications, Keye Tech has launched a defect generation model. Based on defect data and with diffusion models as the technical foundation, it simulates defects in different positions through forward diffusion + reverse diffusion of sample images. This solves problems such as difficulties in model training and verification, and poor model indicators caused by insufficient image data.

Through continuous algorithm optimization, Keye Tech's AI algorithms have evolved from requiring tens of thousands of images to only dozens of images to complete model training. Gradually moving towards lightweight models, it has achieved the optimal balance between speed, power consumption, cost, and deployment convenience.

In the past, Keye Tech's AI model was like a blank slate that needed massive amounts of data to "fill" its knowledge; now, Keye Tech's AI model is an "experienced expert" that can quickly master new tasks with only a small number of samples (40-50 images).

AI Cross-Border Application: More Accurate Innovation

After years of market testing and refinement, Keye Tech's AI algorithms have continuously expanded horizontally, empowering more scenario values.

Industrial Quality Control – Image Matching Algorithm + Stereo Vision

Keye Tech has successively developed image matching algorithms and stereo vision algorithms, realizing AI visual inspection for single-product detection such as printing and electronic components, and improving the yield rate of production lines.

Agricultural Quality Control – 1-Hour Rapid Training

In the field of agricultural quality control, considering the industry environment and time efficiency, Keye Tech took the lead in abandoning the "time-consuming training method" and refined the "1-hour rapid training" mode. It does not require a large amount of image data accumulation; with only 50 images, it can complete data annotation, model training, and model deployment within 1 hour, achieving high efficiency and accuracy.

AI + Robot Arm – 3D Algorithm

In the new direction of AI + robot arm, Keye Tech's AI algorithm measures and reconstructs the 3D geometric information of objects (such as shape, depth, and position) from data. After obtaining accurate 3D information, it can independently identify, locate, and grasp scattered and randomly stacked objects, while also completing defect detection or product classification of objects.