A deep learning approach for plate circular hole detection on composite parts

Abstract

In the aviation field, most parts represented by composite materials have a complex background. How to quickly measure and locate the circular hole of this type of part through visual measurement technology has become a key problem in the industrial field. In view of the difficulties of circular hole segmentation for composite plate parts, such as texture noise interference and the low accuracy of circular hole segmentation. This paper presents a circular hole segmentation method based on deep learning. The U-net network was used to realize the segmentation of the inner and outer textures of the round hole. Then, to solve the problem of missing information in the prediction results of the U-net model, the network model is improved by introducing a dilated convolution and BN layer. The accuracy of the improved model is improved by approximately 2% in the test set. Finally, this paper proposes Zernike moment sub-pixel edge detection of a 7×7 template coefficient to realize the edge sub-pixel location. Finally, the pixel equivalent is solved, and the circular hole contour is fitted using the least-squares method. The measurement accuracy of the improved U-net model in the circular hole of the composite plate ∅(4mm-5mm) reached 0.046mm.