Research On 3D Reconstruction Of Wind Turbine Blade Based On Deep Learning

3D reconstruction of wind turbine blades can provide favorable information for trajectory planning of uav image acquisition,and make defect classification and visualization more clearly and intuitively in the wind turbine blade defect detection system.The traditional multi-view 3D reconstruction algorithm is difficult to recover the 3D information of the weak texture area of wind turbine blades.In contrast,deep learning can better adapt to the reconstruction of the difficult scenes such as weak texture and non-diffuse reflection,and the extracted features have stronger semantic properties.In this paper,3D reconstruction of wind turbine blades is carried out using the multi-view 3D reconstruction method based on deep learning.The main research contents include:Aiming at the weak texture of wind turbine blades and the influence of actual noise in wind turbine blade pictures,SIFT algorithm with rotation and scale invariance characteristics was improved.Firstly,gaussian difference scale space detection method is adopted to detect key points and generate local gradient histogram.The linear direction with the greatest dispersion in the projection space is taken as the main direction of feature points.Then,the coordinate projection process from standard coordinates to polar space is realized,the feature descriptor is established in combination with the distribution information,and the matching is carried out according to the Euclidian minimum vector distance.Finally,the random sampling consistency is adopted to eliminate the mismatching.Experiments show that the algorithm in this paper greatly improves the running speed and the running efficiency of the whole 3D reconstruction while maintaining good accuracy.In view of traditional visual stereo vision algorithm is difficult to deal with the issue of weak texture wind-power blades,first of all,to extract the feature points of sparse reconstruction using SFM algorithm to obtain the camera parameters and the scene information,and then introduced a kind of used for stereo vision MVSNet,multiple points of view to match the price based on 3D convolution network regularization,the depth of the residual network to optimize prediction results;Finally,to solve the problem that 3D convolution regularization is very memory-consuming,GRU regularization is used to reduce memory consumption and obtain a larger depth range,making large-scale scene reconstruction of wind turbine blades possible.Based on multi-view 3D reconstruction OF DTU data set,the network model was trained,the wind turbine blade data set was used for model testing,and compared with MVSNet and open source software based on traditional algorithms for analysis.Experiments show that this method is superior to traditional methods in both completeness and efficiency.Compared with MVSNet,this method can obtain a larger depth range due to the reduction of memory consumption.Through the experimental analysis of the effect of the number of matching views on the performance of 3D reconstruction,it is found that increasing the number of matching views can improve the performance of 3D reconstruction.