Research On Quality Classification Of Guzheng Surface Plate Based On Deep Transfer Learning

At present,the quality classification of guzheng noodles mainly depends on the manual appraisal by instrument technicians,but this method has personal subjective factors,low judgment efficiency and high judgment cost.Since this artificial method of appraising wood grades can no longer meet the needs of the large-scale market,how to realize the scientific and engineering grading work of guzheng panel panels is an urgent issue to be solved.Because the wood cross-section contains the key features of wood recognition such as the size,shape,and duct distribution pattern of the pipe hole,the deep transfer learning technology can efficiently extract these key features from the image.The quality grading was studied.On the one hand,this paper proposes an automatic recognition method for guzheng panels based on deep transfer learning.First,a high-power microscope was used to sample three different grades of paulownia wood to construct a wood duct image data set.Then,the data in the training set uses multiple data enhancement methods such as flipping,rotation,cropping,deformation,and scaling to expand the data.Finally,the deep learning models that have been pre-trained on open source datasets such as ImageNet are migrated to this experiment,and these models are appropriately adjusted and improved to realize the automatic classification and recognition of catheter images.It is worth noting that in the experiment,with the help of visualizing the output feature maps extracted by different convolutional layers,a method of retraining the model by fine-tuning the weight of the last convolution block in the model is determined,so that the model can be initially extracted for the new data set Different features in the image.Compared with the current mainstream methods,it proves the effectiveness of deep transfer learning in the quality classification of guzheng noodles.The recognition accuracy of the fine-tuned Resnet50 model reaches 89.51%.On the other hand,because the research object of this paper is spread between different qualities of the same tree species,the similarity between the images is relatively high.In order to improve the poor effect and high cost of fine-tuning the pre-training model,this paper proposes a method based on secondary features.The extracted deep transfer learning model.Resnet50 is selected as the backbone network,and the network structures such as residual connections,deep separable convolutions and SE modules are integrated in the back-end to build Resnet50-X and Resnet50-SE,among which the deep separable convolution used in Resnet50-X It can increase the accuracy as little as possible while increasing the model parameters,and the recognition speed is faster.Resnet50-SE can realize the feature reuse and the weight distribution of each channel.It has more excellent feature extraction capabilities,and the recognition accuracy reaches 94.39%.The Kappa coefficient reached 91.49%.In order to characterize the performance of the improved model,we compared it with the mainstream deep learning model and found that the method proposed in this paper has higher model recognition accuracy and lower model operation cost.The above two aspects of research are conducive to the realization of high-efficiency quality identification of guzheng surface plates.Among them,the automatic classification and recognition of paulownia duct images is realized based on the deep migration learning technology.In addition,in order to effectively extract the difference features between the paulownia duct images,it is proposed The deep transfer learning model based on secondary feature extraction can obtain higher recognition accuracy while training the model at a lower cost.The above research is helpful to realize the rapid and accurate identification of guzheng surface plates,so as to provide corresponding technical support for the selection of materials and plate identification methods for the broad musical instrument market.