Deep Learning Based Approach For Circular Hole Detection On Composite Parts

Recently years,the use of composite materials in aerospace industry has been increasing due to its performance advantage.The rich texture information on the composite surface poses a challenge for visual measurement,and detection of circular holes on composite parts becomes one of the key issues in vision measurement of robotic drilling system.The traditional circle detection algorithm can not meet the circular hole measurement requirements of composite materials.The difficulty of vision measurement of composite material lies in its complex texture information.The traditional circle detection algorithm has a low recognition rate in the circle detection problem of composite materials,which will seriously reduce the efficiency of robotic drilling.Aiming at the above prob-lems and challenges,this paper studies the high-precision composite circle detection method based on texture segmentation,deep learning and semi-supervised learning.The main innovations are:(1)For the complex texture information of composite materials,an improved LBP(Local Bi-nary Patterns)based texture segmentation method——LEP(Local Exponential Patterns)is pro-posed.The LEP algorithm discards the too many binary patterns of LBP and gives meaningness to the binary value.The LEP value of the central pixel not only reflects the texture information of the area around the pixel,but also the number of textures in the area.The LEP algorithm makes pixels with similar texture information have the same LEP value,is very suitable for the segmentation of texture images,and has gray invariance and rotation invariance.(2)Based on the LEP algorithm,a round hole detection method based on texture segmentation is proposed.Based on the texture segmentation algorithm,we designed an efficient circular hole detection process for composite materials.Firstly,the contour is extracted from the texture seg-mentation result,matched with the circular template,and the remaining contour is elliptically fitted to obtain the candidate circle.Finally,the active contour model is used to optimize the elliptical contour and re-fit to obtain the final circular contour.The method greatly improves the recognition rate and accuracy of the composite hole detection.(3)Aiming at the problem that the accuracy of traditional segmentation algorithm is not high enough and the cost of data labeling process is too large,a method for detecting composite holes based on semi-supervised deep learning is proposed.The detection result of the circular hole detec-tion method based on texture segmentation is used as the pseudo-label of the data,and is optimized by the morphological method.Then the iterative self-training method is used to train the circular segmentation model and optimize the pseudo-label at the same time.The streamlined and efficient deep network is proposed to solve the specific task for the circle segmentation,and a roundness loss function which can represent the circularity of the unknown region is proposed.Finally,a model fu-sion method based on confrontation training is adopted,and multiple training models are integrated into one model.This method significantly improves the accuracy and robustness of composite hole segmentation.