Research On Batch Strategy ELM Image Classification Algorithm

With the advent of the information age,various data,such as text,voice,and image information,have increased exponentially.At the same time,with the rapid development of technology,compared with other transmission media,the sudden increase of image information is more and more obvious,and people's perception of image information is also becoming more efficient.Therefore,the effective capture and classification of image information in the era of big data has become the most basic problem of machine learning.Image classification technology is widely used in industry,agriculture,commerce and other fields,which greatly facilitates people's production and life,and its classification method has become an important research direction.Among many image classification methods,the Extreme Learning Machine(ELM),as a special neural network that does not need to adjust hidden neurons,has the characteristics of fast learning speed,easy implementation,and less manual intervention.Bag of Words model(Bo W)is a feature extraction method derived from text classification.It can better complete image representation.And it is currently a commonly used feature extraction method in image classification tasks.Based on this,this article proposes a batch strategy,and uses it as the main line to propose two ELM variants,aiming to process high-dimensional image data.At the same time,based on the bag of words model and feature coding,the Spatial Pyramid Matching(SPM)model is deeply studied.In the decision-making stage,the Batch inherited Extreme Learning Machine(B-ELM)proposed in this paper is used as the classifier to construct a scene image classification method.The specific research work of this paper is as follows:In view of the high computational complexity and huge memory requirements of ELM when processing high-dimensional data,the high-dimensional data set is equally divided into different batches that can be independently optimized and solved,and the automatic encoder network is used for batch processing to reduce the data dimension;In order to effectively integrate the output weights of each batch,the inheritance factor is proposed,the hidden layer weight connection of each batch is established,and the learning characteristics of each batch are inherited to the greatest extent;according to the regularization framework,the problem of constrained optimization is solved and the target Function,the new hidden layer output weight is obtained through the regularized cost function,and a new method of solving the output weight is obtained.Aiming at the problems of high memory energy consumption and poor generalization in ELM processing high-dimensional data,the data set is processed in batches to reduce computational complexity.At the same time,the self-encoding process based on the L1 norm is introduced in the layer-by-layer unsupervised training,and a multi-layer autoencoder network is constructed to perform unsupervised encoding on each batch of data.Combine the manifold regularization framework and introduce inheritance factors,establish the functional relationship between batches,and construct the objective function of manifold classification with inheritance items.Construct a decision-level classifier by deriving a new output weight formula to maintain the integrity of the data,thereby improving the generalization and robustness of the algorithm.Aiming at the problem of poor feature extraction ability and low classification accuracy of ELM in processing complex scene image classification,the SIFT feature of the image is extracted through the bag-of-words model and quantified into visual vocabulary,and the sparse coding is used to obtain local salient features and reduce the feature dimension;use SPM to generate Multi-level dictionary feature distribution,capture the spatial location information of features,and realize sparse hierarchical representation of image features.In the decision-making classification stage,build a B-ELM model instead of traditional SVM as a classifier,reduce network complexity,and improve the classification of scene images effect.