Quality Assessment of Friction Welding using Image Super-resolution via Deep Convolutional Neural Networks

Abstract

In this research, we devise a deep convolution neural network model for the weld image super-resolution, to determine the trace of the intermetallic compound and also to detect the damages in the welded surface. Firstly, the low-resolution weld image is interpolated using the bi-cubic interpolation approach, which is followed by a non-linear end to end mapping, whose input is a low-resolution weld image and the output is a high-resolution weld image. Also, this approach has a low overhead, since it has a quick pre-processing and post-processing phases, in addition to the optimization of all layers. Experimental results indicate that this approach generates a superior quality image than the weld image produced by means of bi-cubic interpolation. Furthermore, this method surpasses the bi-cubic interpolation approach in terms of Peak Signal-to-noise-ratio (PSNR). The resulting high-resolution weld image aids in the assessment of the quality of the friction welding process