Research On Forage Recognition Method Based On Convolutional Neural Network

China has a vast grassland area,accounting for about 40%of the country's land area,and is one of the countries with the largest number of grassland resources in the world.Among them,forage grass is an important part of grassland resources.It contains a variety of trace elements and vitamins.It is the first choice feed for domestic animals,and as a renewable natural resource,it can be harvested many times a year.Fast and accurate classification and identification of pastures will help realize the digitalization and automated management of grassland pasture resources,and promote the development of animal husbandry and related industries.At present,the classification and identification of pastures mainly rely on manual identification by experts,which is time-consuming and labor-intensive and the results are subjective.Therefore,in order to improve the accuracy and efficiency of forage classification,and the degree of automation of grassland forage digital resource information acquisition,this paper has carried out related research on classification and recognition of forage seed image and forage landscape image classification based on convolutional neural network,the main research content and conclusions as follows:(1)Collect and establish forage grass seed image and forage landscape image data sets.At present,there is no complete and public forage image data set at home and abroad.The laboratory's self-built image database is acquired manually using SLR digital cameras and mobile phone cameras.Collect 10 types of grass forage grass seed images,including old Miscanthus,Elymus cylindrica,Elymus sibiricum,Elymus glabra,Mongol icegrass,Agropyron sibiricum,G.sibiricum,G.sibirica,Ciliated goose watching grass,goose watching grass.The collection environment is a fixed place under a uniform light source,and the black background is uniformly adopted.The image preprocessing process includes geometric normalization of the original image,removing redundant information and segmenting the region of interest.In order to ensure the sufficient training parameters of the convolutional neural network,the seed image data was expanded by methods such as radiation transformation and mapping transformation to establish the forage seed image data set.The pasture landscape image library contains 5 types of common pastures,such as bromegrass,light spike grass,Mongolian grass,old straw,legume alfalfa.The different types of mobile phones and SLR cameras were used to shoot in different weather conditions and different growth periods of forage grass,which increased the diversity of data and provided a basis for verifying the robustness of the model Identification has important research significance.(2)Two classification and recognition methods for grass seed based on convolutional neural network are proposed.Method 1:Through self-establishing multi-layer convolutional neural network model,extract the depth features of forage seed images,and train and recognize by Softmax classifier,the final recognition rate can reach 91.67%.Compared with other traditional recognition methods,such as KNN,Naive Bayes,SVM,etc.,based on LBP texture features,the recognition rate is 7%-44%higher on average,and the model is very robust,The fitting effect is good during training and testing.Method 2:Improve the VGG-16 model in VGGNet to a VHS model for classification and identification of forage seed images.Through the transfer learning method of feature vector fusion and fine-tuning parameters,the model is trained and tested under different data sets and parameter setting benchmarks,and 36 sets of recognition results are obtained.The final accuracy rate can reach up to 94.14%.Under the relatively optimal parameter setting,the model can quickly reach the optimal solution and the fitting effect is good.(3)Proposed classification and recognition of forage landscape images based on convolutional neural network and hinge regularization loss function.The original CNN model used the cross-entropy loss function and Softmax classifier.Under this method,the model cannot adapt to a high learning rate,and it is prone to serious parameter overflows that lead to abnormal training termination.In order to improve the training speed of the model,and at the same time ensure the good robustness and classification accuracy of the model on the data set at a high learning rate,this paper first proposes a hinge regularization loss function,and then builds multiple SVM classifiers under a multi-classification strategy,and With reference to the structure of VGGNet convolutional neural network feature extraction layer,a VHN forage landscape image classification model is established.Finally,the classification accuracy of the VHN model under the transfer learning method and high learning rate(maximum 3)reaches 95%,which effectively improves the model training rate and has good robustness to the pasture landscape image data set.At the same time,under the same learning rate,the training rate of this method is higher than the cross-entropy loss function.