Research On Board Defect Image Classification Based On RegNet

Board defect detection is one of the important processes in board processing.The realization of automatic detection of wood board defects is a prerequisite for high-quality board processing,and the recognition of board defect images is the difficulty of this technology.Preferred saws do not fully implement automated technology,and in the production process,it is still necessary to manually line defects in wood panels.In order to alleviate manual pressure,this study intends to replace workers with cameras and computers to identify and classify defects on the panels.The recognition accuracy of board defects based on traditional algorithms is not high for large sample data.As the design mode of network models changes from manual exploration to automatic search,many automatic search network structure models are widely used in the field of image recognition.However,currently,in the field of classification research on board defect images,the main algorithm is still the convolutional neural network designed manually.In order to apply the automatic search network to the field of board defect image recognition and classification,this paper proposes a relevant model based on Reg Net design space.The specific research content is as follows:In view of the phenomenon that the accuracy of traditional algorithms to identify board defect images is not high,this paper studies the board defect classification model based on Reg Net design space.First,the small sample database is expanded into a data set through the data enhancement method of supervised single sample type;Then put the sheet defect data set into the Reg Net design space to verify the feasibility of the Reg Net automatic search space for image classification.In order to further improve the recognition accuracy of the model and reduce the model parameters,this paper proposes a board defect classification model based on attention mechanism and Ghostconv structure,and embeds multiple attention mechanisms into the block module of the bottomleneck part of the Reg Net design space.Through experimental comparison,the attention mechanism module that is most suitable for training the board defect data set is obtained.On this basis,the Ghostconv structure is introduced into the block module to replace part of the group convolution structure in this module,reduce the model parameters and improve the image recognition rate.From the experimental results,it can be seen that the model has a high classification accuracy for the data set of board defect images.In view of the problem that the accuracy function of the verification set rises slowly in the board defect classification model based on the attention mechanism and Ghostconv structure,this paper proposes a board image classification model based on multi-scale feature fusion.First,based on the Ghost Reg Net+NAM model,an improved Bi FPN feature fusion model is added to increase the feature extraction capability of the middle layer.Then the loss function of feature extraction is adjusted to multi-classification Focal Loss loss function to improve the extraction ability of feature information.The experimental results show that the model is more stable for the recognition and classification of board defect images.