Research On The Compression Method Of Deep Learning Models For Bamboo Image Classification

Bamboos are the general name of Gramineae bamboos subfamily.There are about 88 genera and more than 1400 species of bamboos in the world.The area of bamboo forest is more than 37 million hectares,mainly distributed in Asia Pacific,Latin America and Africa.As an important multi-purpose forest resource,bamboo plays an important role in the regional development and social development of the world,and its research has been paid more and more attention all over the world.The recognition of bamboo species is not only an important premise for bamboo scientists to study bamboo properties and applications,but also a way for people to understand bamboo and nature.In recent years,with the significant progress of deep learning in the field of image classification,it is possible for bamboo image classification and recognition.However,due to its large number of parameters,high redundant weight and huge consumption of computing resources,it is difficult to apply it to the embedded platform.For this reason,this paper focuses on three problems of bamboo image data collected under natural light: establishing a deep learning bamboo species classification model,pruning and quantifying the established deep learning model,transplanting the compressed model to the mobile phone,and carrying out off-line bamboo species classification.The main research work of this paper is as follows:(1)The bamboo species classification model based on deep learning was studied.In order to better extract the characteristics of bamboo species and reduce the influence of irrelevant variables on the classification effect,the deep learning algorithm is used to classify bamboo species.The convolution neural network is used to abstract the original image hierarchically,and the VGG-16 network structure and Alex Net network structure are used to train the classification model of bamboo species.(2)Based on the "pruning training" method,the convolution neural network of bamboo species classification is compressed.On the premise of maintaining the accuracy of neural network,VGG-16 network and Alex Net network structure are pruned.Through the dynamic threshold setting method,redundant connections are deleted,only the effective weights within the threshold are retained,and the storage space required by neural network is reduced.Pruning simultaneously trains the network structure and extracts the features,so as to better maintain the network structure after pruning.The experimental results show that the memory consumption of the model based on VGG-16 network structure and Alex Net network structure is the same,but the accuracy of the test set after pruning is lower than that before pruning,and the accuracy of the test set after pruning of Alex Net network structure is less.(3)The weight sharing method was studied based on K-means + + algorithm.For the pruned model,the weights of each layer of neural network are clustered and quantified by Kmeans + + clustering.The training results show that the memory occupation of quantitative model based on Alex Net network structure is less than that of VGG-16 network structure model,and the bamboo classification model based on Alex Net network structure can be transplanted to the mobile phone better.(4)A bamboo species classification system based on Android was developed.Through Java programming on Android studio platform,the bamboo species classification system is designed and developed,including picture upload module,picture classification module,etc.The compressed bamboo species classification model was transplanted to the server,PC and mobile phone for testing.The effect of the compressed bamboo species classification model under different hardware environment was studied.The effectiveness of the method and technology was verified by an example.The research content of this paper provides technical support for the rapid and convenient classification and recognition of bamboo germplasm resources.The compression of bamboo species model based on deep learning is of great significance to the research of bamboo species image processing.