Research On Image Classification Method Based On Local Receptive Field Extreme Learning Machine

In the era of rapid technological development,the speed of dissemination of various types of information is increasing exponentially,and people are increasingly using computers to deal with various problems.As the core content of machine learning,image recognition and various network technologies,neural networks have been widely studied and applied,and have become one of the important methods to solve various machine learning problems.Among them,image classification,as a basic research problem of machine vision,has gradually become an important research content in the fields of image recognition and pattern recognition.Therefore,the research on the algorithm and structure of image classification is of great significance,and the extreme learning machine based on the local receptive field has attracted more and more scholars' attention with its more efficient training methods and easier-to-understand derivation theories.However,in all kinds of image classification problems,extreme learning machines based on local receptive fields also have problems such as insufficient feature extraction,insufficient ability of single-layer networks to extract deep semantic information,and certain limitations in processing large-size images.Therefore,this article focuses on the study of extreme learning machines based on local receptive fields,and optimizes and improves this network model to improve the image classification performance of the network.The specific research work of this paper is as follows:First,in order to solve the problem of insufficient image feature information extracted by the extreme learning machine based on the local receptive field,multiple kernel empirical learning network is constructed.The network progressively enhances the original image,and inputs this image and the original image into the network in parallel.The local receptive field is used to extract the detailed features of the image,and the global receptive field trained by the extreme learning machine auto-encoder is used to extract the global contour feature,so as to make full use of the local and global feature information in the image.In the feature classification,the low-rank matrix mapped by the multiple kernel empirical is used as the hidden layer of the network to solve the network output weights,to achieve the complementary advantages of each part of the feature,and to verify the training efficiency and classification ability of the network through experiments.Secondly,in view of the insufficient number of network channels and the inability of a single-layer network to extract more color information and deep-level feature information from the image,three-channel receptive field empirical kernel extreme learning machine based on cosine similarity is proposed.The original image is separated by RGB three channels,and the local and global receptive fields are used to perform multi-layer convolution pooling on the three-channel image.The cosine similarity measure is used to calculate the cosine similarity between features,and this similarity matrix is used as the information matrix for solving the network output weights,making full use of the associated information between the features.The method of empirical kernel mapping reduces the feature dimensionality,alleviates the impact of high-dimensional features on network performance,enhances the expression ability of features,and further improves the overall classification performance of the network.Finally,in order to solve the problem of insufficient classification ability of the proposed improved network when processing large-size images,correntropy extreme learning machine based on spatial pyramid matching and local receptive field is proposed.The network introduces a spatial pyramid model to extract more feature spatial distribution information.Combine SPM features with local receptive fields,and use receptive fields to extract features from different levels of feature distribution maps.The network uses the cosine similarity measure to constrain the network weights.On the basis of the traditional training method,it constructs the discriminative constraints of the modulus length and cosine similarity based on the correlation entropy criterion,and derives the correlation entropy update formula to train the network output weights.Experiments prove the effectiveness of the network in large-size image classification tasks.