Transfer Learning Via Identical Distribution In Convolutional Neural Networks

With the development of deep learning,various application scenarios have emerged.How to apply excellent convolutional neural network models or algorithms to actual scenes has become a common problem.First,networks with excellent performance,while performing very well in standard big data sets,are difficult to apply directly to complex real-world environments.Second,collecting large amounts of data in complex application environments requires significant time,labor,and equipment costs.The quality of the model algorithm is not only limited by the amount of data,but also hardware limitations such as device mobilization in the application scenario.At the same time,the computing power of these mobile devices is quite limited,and the computational realtime performance of complex convolutional neural networks cannot be guaranteed.The storage space of the devices also does not allow for a large number of complex network nodes.Combining the above-mentioned related issues explained in the first chapter,this paper proposes a comprehensive solution in Chapters 2 and 3.Try to apply the knowledge in the source dataset to the target scenario by establishing a identical distribution constraint to help the target scenario build the model.In addition,in the process of model knowledge migration,the framework solves the problem of lack of data and data in the target environment.The entire solution is based on the identical distribution of feature space distribution between convolutional neural networks.This method establishes a migration learning framework by fitting the performance of complex networks with a miniaturized convolutional neural network.It is also used in a variety of standard data sets.In the repeated demonstration and practical use of multiple experiments,the knowledge migration framework based on the identical distribution constraint of convolutional neural networks is fully validated.We use the proposed framework to solve several problems in real-world applications.In the fourth chapter,such as the near-infrared face recognition algorithm,the face-to-face comparison of the face and the object of the server,and other practical problems.At the same time,we prove that the framework has the characteristics of fast training and fast convergence.In Chapter 5,we added the design and demonstration of incremental object recognition under the framework of distributed consistency constraints.The work of this thesis is carried out in three aspects: firstly,the feasibility and validity of the identical distributed constraint framework are proved by the argument of theory and basic experiment Then the near-infrared face is solved in detail through many field problems of face recognition.Identifying the difficult points in the field of comparison with human witnesses.Finally,by extending the identical distributed constraint framework,it is applied to the object-seeking task of the service robot,and obtains an acceptable recognition rate under the constraints of minimal models in a small number of categories.