Research On Pattern Classification And Image Recognition Technology On Hyperspherical Crown

With the gradual development of computer technology and artificial intelligence,pattern classification and image recognition technologies have been widely used in various fields,such as science,medicine and economy.Different pattern classification tasks and scenarios often require different classification techniques and models.For the classification tasks of some standard datasets(For example,the UCI data for machine learning proposed by the University of California,whose attribute values and categories have been labeled),feature extraction and classifier construction are mainly considered.Due to image is susceptible to noise pollution in the process of acquisition and storage and the image dimensions are high,the three main aspects of noise preprocessing,feature extraction and classifier construction are mainly considered for the recognition task of image datasets.Each aspect has a significant impact on the final image recognition result.Although the existing noise removal methods and pattern recognition models have made great progress,the current general view is that there is no single model applicable to all pattern recognition problems.We need to combine the corresponding pattern recognition algorithm with the specific problem.Because of the structural complexity and diversity of different datasets,the existing methods have deficiencies in some practical classification.For example,some classification algorithms work well on the noiseless datasets but have unobvious classification performance for the noise datasets;some classification algorithms are not strictly limited by the data distribution form,while others are sensitive to the data distribution form.For the supervised pattern classification problems and transferable image recognition tasks,this paper focuses on optimizing the model structure and improving model recognition ability.It mainly studies the key technologies of image noise removal,feature extraction and classification recognition,and obtains the following innovative research results.1.A method of impulse noise removal in color image preprocessing is proposed.Firstly,this method analyses the characteristics of correlated impulse noise and uncorrelated impulse noise in the noise sensing model of image signals.The signal-to-noise ratio difference between the two kinds of impulse noise in the vector image and component image is analyzed by using signal detection theory.Then a fuzzy detection technology is performed to detect noise on the pixels of each channel.Based on the noise detection results,the pollution degrees of the center pixel and the neighborhood pixels are analysed,and two newly defined pixel subsets are used to classify the pollution degree of different pixels.Thus three new subfilters are proposed accordingly: a similarity weighted mean filter within the partial relation noise-free neighborhood subset,a vector filter within the noise-free neighborhood subset and a noise-free component spatial distance weighted mean filter.The experimental results of different types of images show that the multi-stage switching hybrid filtering method can automatically detect the difference of pollution degree in different regions.It can make full use of the local correlation between neighborhood pixels and the correlation among different channels to filter impulse noise and protect image details.2.A discriminant analysis algorithm based on hyperspherical crown mapping in feature space is proposed.For the separability problem casued by heterogeneous sample fusion in the hypersphere mapping,a unit hyperspherical crown model in feature space is proposed.It has many important properties which are superior to the original hypersphere mapping.These properties mainly include keeping the structure information of pattern vectors unchanged in the nonlinear mapping process,keeping the distance order relationship between the mapped vectors unchanged,and improving the separability of the sample set under certain constraints.Based on the properties of the proposed model and the nature of the scatter matrix,an inner product nearest neighbor classifier on the hyperspherical crown manifold is constructed and the corresponding incremental learning method is analyzed.The research results show that the proposed model can keep the manifold dimension of the feature distribution unchanged.Even the samples overlapped under the traditional normalized mapping are separable under the proposed model.Experimental results verify the effectiveness of the proposed model and discriminant analysis algorithm.3.An optimized probabilistic neural network classifier based on hyperspherical crown mapping and adaptive kernel coverage is proposed.Firstly,an adaptive kernel coverage method based on the global and local information of sample space is built to generate pattern nodes,which simplifies the pattern layer structure of probabilistic neural network.To estimate the probability density of different regions more accurately,the corresponding intra-class weight coefficients are assigned to different pattern nodes,and the network parameters are optimized by using the expectation maximization algorithm and Lagrange algorithm,so the probability density estimation of the optimized network can be closer to the real probability density.The results show that the classification surface of the proposed network approximates the optimal interface under the Bayesian criterion.It can automatically adjust the number of pattern nodes and the corresponding network parameters,and simplify the network structure.Experimental results in different datasets show the proposed algorithm has good robustness and classification ability.4.A new hyperspherical crown probabilistic convolution neural network for transferable image recognition is proposed.Firstly,the relationship between the generalization ability of the convolutional layer and the class number of training set is studied from the aspects of category increase,category reduction and unknown category expansion.The proposed network makes full use of the generalization ability of image feature extracted by the convolutional layer and the transferable ability of model.It mainly consists of three parts: the training phase of feature mapper,the network parameter conversion and the training of adaptive classifier.The research results show the feature mapper trained in the multi-category dataset has better generalization ability and the BP classifier in the convolutional neural network is not always the optimal classifier.The proposed deep hybrid network does not need training or fine tune on the target set,which improves the efficiency of network learning and saves the training cost.The experimental results on different datasets verify that the proposed deep hybrid network has good classification ability and stability.