Applicative Framework for End-to-End AOI Implementation
1 . Identification of rotors inside the picture . The used algorithm first recognizes the location of each rotor through unsupervised semantic segmentation based on pixels clustering . Pixels are singularly classified and then grouped through a convolutional neural network , trained by minimizing a loss function which takes into account both pixel value and their spatial arrangement . The idea relies on the fact that pixels with similar values and spatially close belong to the same object . The rotor class is identified , and the output consists of the coordinates of each rotor location , which are used for their extraction from the original image .
2 . Recognition of the rotor typology . A second algorithm spots the smaller circles identifying the different classes . Each rotor detected in the previous step is isolated and processed singularly through the Hough Circle Transform method , parametrized to spot the circumferences on the rotor metal coupling and , as a consequence , the correspondent class .
The outcome of the AI algorithm is compared to the recipe set on the PLC , thus verifying if all rotors inside the box belong to the right class . Finally , a visual outcome of the control is displayed on the PC monitor by overlaying red marks on the original image at the positions where the misplaced rotors are detected .
Figure 3-3 : Final processed image ( left images ) shown on screen after the AI algorithm analysis is performed during production of Class C , indicating wrong rotor type belonging to class A with a red cross .
12 March 2022