VIDEO INSPECTION SYSTEM FOR COLD-ROLLED STEEL ROLLED PRODUCTS
SKU0072
р.
р.
DESCRIPTION
The system for detecting and classifying surface defects in rolled metal products integrates two types of illumination - diffuse and side illumination - to provide analysis of defects with different structures. The use of color high-speed cameras with high contrast (Weber - not less than 0.25) makes it possible to detect small defects up to 0.5 x 0.5 mm. Based on the Faster R-CNN convolutional neural network, a model optimized for real-time application has been created, which combines several classifiers tuned for different lighting conditions. This improves the recognition accuracy to under- and over-reject rates of no more than 3% and 5%, respectively, and up to 10% and 15% when defect classes are taken into account. The image preprocessing method improves the quality of optical inspection by calculating the statistical characteristics of each image and subtracting the average intensity value to equalize the illumination. This approach increases the contrast of the defect image and minimizes distortion from uneven illumination. In addition, the system saves the processed images for further analysis and training of neural networks.
ADVANTAGES OF THE DEVELOPMENT
Increase of speed and accuracy of rolled products certification
Automating the certification process without sacrificing quality
Import substitution and independence from foreign suppliers
Support for different lighting conditions for optimal certification
Reduction of certification costs through automation
Easy integration and scalability for different types of rolled products
Adaptive training and system modernization
The system provides high attestation accuracy and stable operation under industrial conditions
CHARACTERISTICS
Cameras: High-speed color (4K, ≥1000 fps)
Lighting: Combined (ambient and side illumination), automatically adjusted for accurate detection
Algorithms: Pre-processing for contrast enhancement and detection of defects from 0.5×0.5mm, Faster R-CNN neural network for recognition and classification
Accuracy: Undersampling - up to 3%, oversampling - up to 5%
Performance: Processing time down to 0.01 sec/image