Resumen de Deep Learning Based Algorithms for Welding Edge Points Detection

Ander Muniategui Merino, Jon Ander del Barrio, Xabier Zurutuza, Xabier Angulo, Iñaki Silanes, Uxue Irastorza, Aitor García de la Yedra Basurto, Ramón Moreno

• Low defective rates and high variability in industrial processes, make difficult to develop accurate, reliable and fast automatic quality control systems. This paper presents two novel methods for the Quality Control of welded parts based on computer vision. The first method, BG, uses a modification of the bisection method, and a tuned version of the GoogleNet. The second method, BGN, is an ensemble of GoogleNet CNNs and a Convolutional Auto-Encoder to absorbe observed data variability. The CAE is used to select the best CNN from the ensemble to use for each input image. Both methods have been tested on more than 105 images and run in less than 0.2 s in a standard i7 CPU with mae values around 0.04–3.63 pixels, standard deviation of absolute errors around 1.15–7.48 pixels, and a percentage of correct predictions between 95–99.97%.