Research On Intelligent Detection Method Of Defect And Grain Orientation For Optimal Processing Of Wood Beams And Columns

Wood structure beams and columns are widely used in products such as wooden sunshine rooms,wooden houses,wooden doors and windows,and wooden curtain walls.Due to the existence of defects such as knots,cracks and wormholes in solid wood beams and columns,their appearance quality and mechanical properties are seriously affected.In addition,when the grain direction of wooden beams and columns is opposite to the direction of cutting movement,it is easy to cause quality problems such as burrs,and even splitting.However,due to the large randomness of defect types,location distribution,and size,and texture features have their natural attributes,accurate and rapid detection of defects and grain directions has become one of the difficulties in the research of optimal processing of wooden beams and columns.Existing wood processing equipment has many problems such as low intelligence,low detection efficiency,and low wood yield rate.The organic combination of artificial intelligence technology with wooden beam and column defect detection,grain direction recognition,and layout optimization is an important way to improve wood quality.The yield rate of beams and columns and the important way to ensure their mechanical properties have important theoretical significance and application value for realizing the optimal processing,improving the intelligent level of wood processing and production efficiency.To solve the problems of insufficient samples of wooden beam and column defects and poor network robustness,the self-made COCO（Common Objects in Context）format data set was enhanced by online data enhancement,Grid Mask data enhancement and Mix Up data enhancement methods to enhance the generalization of the model and robustness.The Cycle GAN（Cycle Generative Adversarial Network）network is used to generate new defect sample images to solve the problem of uneven wood defect categories.Select wood texture pictures without defects from the collected images,crop them,randomly rotate them,and finally get the wood texture data set.Introduce the principle of heat map to improve the label of knot data set,and establish knot medullary data set.The data set preparation provides data support for the subsequent research work of defect detection,texture recognition and pith location.Aiming at the problems of poor generalization ability,complex model,large amount of parameter calculation,and poor real-time performance in the existing defect detection based on deep learning,a wooden beam and column defect detection algorithm based on the improved YOLOv3（You Only Look Once v3）model is proposed.Use the lightweight model Mobile Net V3to replace the backbone network of the original YOLOv3,and replace the activation function with H-Swish.Compared with the original YOLOv3 network,the average precision（AP）of the designed model has increased by 6%,AP₅₀ has reached 92.4%,the number of parameters has been reduced by 62.35%,and the prediction speed has reached 53.5FPS.In terms of defect segmentation,in order to solve the problem that the Mask R-CNN residual network has insufficient ability to extract context information,a layered residual connection module is established to extract the feature information of multiple receptive fields for each layer of the network;For the problem of insufficient fitting ability,the improvement of the deformable convolution method is adopted to improve the adaptability of the model to irregular geometric deformation defects.Compared with the original model,the detection accuracy of the improved algorithm is increased by 9.9%,and the segmentation accuracy is increased by 5.7%.In order to ensure the processing along the grain in the processing of wooden beams and columns,a small-sample wood grain direction recognition method based on the improved meta-learning Reptile algorithm was proposed to solve the problems of high environmental requirements and poor robustness of wood texture image acquisition in traditional image processing methods.Improve the model calibration part of the traditional Reptile algorithm,remove the classification layer of the last layer of the neural network,and randomly divide the data of the support set S in D_train into two groups,use the high-dimensional space feature vector expression of these two groups of data and the self-learning cosine similarity measurement method further fine-tunes the network so that the input data can be more fully learned.The experimental results show that the improved Reptile algorithm has achieved an accuracy rate of86.6%in the recognition of wood grain direction in the case of 10 shots,and the recognition accuracy of 100 shots has reached 92.7%,and the accuracy of the algorithm is in the cases of 1shot,5 shots,10 shots and 100 shots.Compared with traditional computer vision,ordinary deep learning,transfer learning and the improved Reptile algorithm,it has all improved.To solve the existing problems of weak convergence ability,single-objective optimization and poor search effect of wood ranking sample optimization based on traditional genetic algorithm,an improved multi-objective wood beam column ranking sample optimization method based on NSGA-II（Non-Dominated Sorting Genetic Algorithm-II）algorithm is proposed.In order to solve the problem of insufficient population diversity in genetic algorithm,reverse population is generated based on reverse learning to increase the search ability of the algorithm.To solve the problem of small changes in individual differences in algorithm evolution,a combination of directional variation and uniform variation is used to improve the optimization effect.Taking the actual production problem as an example,the improved NSGA-II algorithm is used to optimize the wood yield and value as the optimization objectives for the wood beam column row sample preferences.The experimental results show that the improved multi-objective optimization algorithm has better optimization effect and stability than the NSGA-Ⅱalgorithm.The number of convergence iterations is lower than the traditional genetic algorithm,the error rate is reduced and multi-objective optimization can be achieved simultaneously.Finally,the Qt language is adopted to design the human-computer interaction system for wood structure beam and column optimization processing with functions of defect location identification,intelligent layout,and grain orientation identification.