Image Classification Based On Multi-instance Multi-label Learning

With the rapid development of Internet technology and information technology in today's society,and the popularization of multimedia devices such as images,the amount of information in image data has increased dramatically.Therefore,how to classify these images effectively and accurately has become a technical problem and research hotspot to be solved urgently.The main work of this article is as follows:First of all,for the traditional distance measurement method can not accurately measure the distance between the package and the package,this paper proposes a self-adjusted distance measurement method,this method can automatically adjust the distance according to the characteristics of different data sets,making the distance measurement more accurate.And proposed a multi-example multi-label image classification algorithm based on self-adjusted distance metrics.Secondly,in the case of unbalanced sample sets,the multi-instance multi-label radial basis function neural network produces an imbalance in the number of hidden neurons,and ignores the class with fewer samples during training,making the classification effect worse.To solve this problem,this paper proposes a radial basis function neural network structure optimization algorithm for image classification.In the first stage,the algorithm optimizes the number of neurons and the center and width of each class according to the distribution of training sets,so that the neurons on each class in the unbalanced sample set are in balance,which reduces the impact of unbalanced samples on network performance.In the natural scene image data set established by Zhou and other scholars,the two proposed improved algorithms are compared with classical image classification algorithms.The experimental results verify the effectiveness and superiority of the proposed algorithm.