Carbon Fiber-Reinforced Polymer Composites Texture Angle Regression Based on the Improved Deep Hough Network

Carbon fiber-reinforced polymer (CFRP) composites have been successfully used in the fields of high-end manufacturing industry because of its merits in high stiffness and resistance. The composites fiber angle is an important factor in the composites overall mechanical properties (e.g., the deviation of 5
$^\circ$
could reduce the mechanical stiffness by 20%). However, the state-of-the-art fiber angle estimation methods have two inherent problems, i.e., potential failures in detecting the composites region of interests (ROIs) and the inaccurate regression in the presence of noises. To solve these issues, this article designs an improved deep Hough network with channel attention residual fusion UNet aggregation (CU-IDHN). It novelly merges the channel attention residual fusion model into the U-net segmentation network to discriminate the composites ROIs from the background. In addition, an improved deep Hough network is designed to select the correct texture candidates and regress the texture angle in a coarse-to-fine manner. The proposed method has been extensively evaluated on the collected composites fiber texture datasets (CFTDs). The results show the competitive performance against the state-of-the-art in both ROI segmentation and CFRP texture angle regression.