| Deep learning is being more and more extensive used in image identification with the development of technology.Based on machine learning algorithms,it can achieve extremely high accuracy and has achieved good practical results.Fine-grained Image Classification is a kind of more detailed classification,it is more difficult to make a accurate classification because different subclasses of the same parent class have a higher similarity,and dog breeds recognition is a kind of Fine-grained Image Classification.As raising pet rate higher,pets,especially pet dogs occupy an increasing proportion in human life.Globally,there are more than 300 breeds of dogs with different appearances.Most people can only recognize some common breeds,and it is difficult to distinguish rare dogs.In fact,each kind of dog has its own unique characteristics.Through Fine-grained Image Classification and the appearance characteristics of different parts,the breed of dog can be judged,and it can provide help for keeping pet dogs and animal management.In order to promote the harmonious symbiosis between humans and dogs,help humans understand the characteristics of different dogs,and carry out scientific feeding.This article mainly carried out the following work:(1)Design,train and optimize dog breed recognition algorithms.Designed dog breed recognition algorithm,and made a classification for 120 common dog breeds.By comparing three traditional Convolutional Neural Network with three Light-Weight Convolutional Neural Network,integrated more effective ones together and improved the accuracy to 82%.Improved the recognition response speed by using a new lightweight model,and applies transfer learning methods Reusing existing models and common features greatly reduces the time cost,data and hardware resource overhead of training models,and solves the problem of model overfitting.On this base,this thesis built a new data set for 50 common dog breeds in China,each has at least 70 high quality images.It can better adapt to the needs of Chinese dog recognition,making the final model accuracy more than 90%.(2)Carry on the overall design and module division to the dog management system.This system is a pet dog management system with dog recognition technology as the core,mainly including login module,image upload module,dog breed recognition module,dog data management module,dog photo module,feeding suggestion module,database storage module,request interface module,process monitoring module,covering most scenes of daily pet raising.The system function and performance requirements are specified,and related interfaces and database structures are designed.(3)Implement and test the dog recognition management system.Application of convolutional neural network,web front-end technology,node.js technology,database technology to realize the dog management system,Completed the various functional modules of the front and back ends,joint debugging and go online successfully.Through the uni-app framework,the dog recognition management system can be operated on the browser web based,browser mobile based,WeChat mini-app,and native APPs.Complete server-side interface development,Established local,test and online environment database,completed cloud server deployment,Tested the system and sorted out the test results.The WeChat mini-program version of the system with optimized model has been completed online,and good results have been achieved.In the future,the existing data set should be further expanded,the accuracy of the model should be improved,and the system experience should be optimized. |
