Research On Cross-domain Object Recognation Optimization Alogrithms Under The Unsupervised Conditions

Multi-scene and cross-domain applications have become a new trend in the field of computer vision.Since the application scene is very complex,it is difficult to collect enough training samples for each new domain.It is necessary for the object recognition algorithm to reuse a small number of label samples or unlabeled samples and extended the recognition model to new application fields.This thesis focuses on cross-domain object recognition under unsupervised conditions.The major purpose of this thesis is to overcome the limitations of current object recognition algorithms in cross-domain applications,and to promote the cross-domain generality and recognition performance.The research work and main contributions of this thesis are as follows:Due to the lack of label information in unsupervised samples,the conditional probability distribution model of samples can not be established accurately.The traditional cross domain recognition algorithm based on probability distribution adaptation can not be effectively implemented.Therefore,a unsupervised probability distribution modeling method is studied,which utilized the prior information of label distribution and pseudo label prediction algorithm to establish the conditional probability model more accurately.On this basis,a joint distribution adaptation algorithm is proposed,and the cross domain generalization ability and error boundary of the algorithm is studied based on PAC learning theory.The algorithm can correct the difference of marginal and conditional distribution simultaneously,so it has better performance.The experimental results show that the algorithm is effective under unsupervised condition.When the samples are noisy,the recognition rate will decrease due to the deviation of data distribution modeling.Therefore,a cross domain object recognition under the condition of noise is researched.Combined with the theory of probability distribution adaptation and parameter regularization,a model parameter similarity regularization cross different domain is derived,and the data distribution modeling method based on the regularization is studied.On this basis,an improved cross domain object recognition algorithm is proposed by the matrix orthogonal decomposition theory and deep learning technology under noise condition.The algorithm builds data distribution indirectly with the model parameters,reducing the influence of sample noise and improving the recognition accuracy.Experimental results show that the performance of our cross domain recognition is better than that of traditional methods.When the number of unlabeled samples in the target domain is very scarce,it is difficult to extract sample features and establish the corresponding domain distribution model.In order to solve this problem,a cross-domain object recognition algorithm based on Attribute Graph Model(AGM)is proposed.Attribute learning method is used to learn the semantic attributes of objects from a small number of unlabeled samples as an effective representation of sample features.We use deep graph networks to learn the semantic relationship between attribute features,and build the attribute relationship graph model of samples to solve the problem of domain distribution modeling.Based on this model,the distribution differences of different domains are adapted to achieve cross-domain object recognition under the unsupervised few-sample conditions.The experimental results show the superiority of the proposed algorithm under the condition of unsupervised few-sample.Using samples from multi-source domains will provide more abundant information,further improving the accuracy of cross domain recognition.Therefore,the establishment of cross domain object recognition algorithm model under the multi-source condition is studied and applied to the flight object recognition.The problem of multi-source distribution adaptation in flight object recognition is analyzed.The single source distribution adaptation algorithm under the condition of large sample is extended by using multi-branch GAN neural networks.On this basis,a multi-source selective domain adaptation algorithm(MSDA)is proposed.In order to further improve the real-time and feasibility of the algorithm,the algorithm acceleration technology based on multi-core heterogeneous processing platform is studied,and the flight object recognition algorithm based on MSDA is optimized.Finally,the effectiveness and practicability of the algorithm to solve the problem of flight target recognition are verified by simulation experiments.