Research On Image Recognition And Extraction Of Oriented Strand Board Surface Based On Semantic Segmentation

Oriented strand board(OSB,oriented strand board),as a new type of environmentally friendly board,has become one of the important choices for construction,furniture and decorative boards due to its high strength and good stability.The shaving shape and its orientation angle in the slab have a great influence on the physical and mechanical properties of OSB such as static bending strength and elastic modulus.Therefore,in the production process of OSB,the directional paving and detection of shavings is one of important link.Traditional methods relying on manual or laser measurement of directional pavement angles are inefficient and cannot meet the needs of rapid detection on automated production lines.With the development of computer vision and AI chip technology,the use of image segmentation technology based on edge hardware devices provides a better solution for the rapid detection of directional pavement angles.Compared with traditional image segmentation techniques,semantic segmentation algorithms based on deep learning have made great improvements in segmentation accuracy in recent years.These algorithms use deep neural networks to learn feature representations of images,enabling models to better understand semantic information in images.Aiming at the problems of low color discrimination and inconspicuous boundaries between particle sheets,this project proposes an improved Deep Labv3+ semantic segmentation model,and at the same time designs and implements an OSB surface image segmentation recognition system based on Jetson Xavier NX,which has been verified.The average inference speed of each image of the system is 0.62 s,which can meet the needs of rapid detection on the production line.The main work content of this paper is as follows:(1)A complete dataset generation and enhancement scheme is proposed.Use a 3024×4032camera to collect the surface image of the pressed oriented strand board.After preprocessing,use the Labelme tool to annotate the image and convert the label format in batches to generate a data set.At the same time,use a variety of data enhancement methods to expand it.Finally,use Python to randomly divide the training set and verification set,where the number of images in the training set is 13192,and the number of images in the verification set is 2328.(2)A large number of comparative experiments were carried out on OSB surface images using traditional image segmentation algorithms and deep learning semantic segmentation algorithms.Three traditional image segmentation algorithms,Canny edge detection algorithm,marker watershed algorithm,and K-means clustering segmentation algorithm,and four deep learning semantic segmentation models,FCN,PSPNet,Deep Labv3,and Deep Labv3+,were used to test OSB surface images.The experimental results show that the Deep Labv3+ network model has the best effect on image segmentation and recognition of OSB surface.(3)In view of the shortcomings of Deep Labv3+ in the segmentation of small-sized shavings and the details of the shaving boundary,this paper improves the encoder and decoder structures of the Deep Labv3+ model.First,in order to improve the prediction performance of the model,the encoder part replaces the backbone network Xception65 with Res Net50,and the decoder part introduces a dual attention mechanism to enhance the model’s ability to extract edge features.In order to adapt to different shapes of shavings,deformable convolution is introduced to improve the segmentation accuracy.The mixed training of weighted cross-entropy loss function and Lovasz Softmax loss function is used to deal with category imbalance,so as to improve the average intersection and union ratio of the model.The test results of the improved model on the small public dataset mini＿city show that its segmentation accuracy has increased by 6.1%.(4)In order to meet the needs of rapid detection on the directional pavement production line,this paper uses the Paddle Seg tool to transplant and deploy the improved Deep Labv3+ model to Jetson Xavier NX.At the same time,for the convenience of user interaction,a set of oriented strand board surface segmentation recognition system is designed by using Qt.This paper proposes an improved Deep Labv3+ model based on the segmentation and recognition of directional particleboard surface images,and builds an embedded system to deploy the model.The test results show that the system can segment and recognize the shavings image quickly and accurately,and calculate the proportion of different kinds of orientation angles.