Research On Deep Multiple Instance Learning Algorithm

In recent years,Multiple Instance Learning(MIL),as a Weakly Supervised Learning method,has achieved a large number of research achievements and has been widely applied in many fields such as education,security and medical treatment.With the development of Deep Learning,Deep Multiple Instance Learning,which is used for efficient visual information retrieval and classification has attracted extensive attention of researchers,and great progress has been made.However,unknown instance label still brings great challenges to the research of Deep Multiple Instance Learning.In order to further improve the accuracy of Deep Multiple Instance classification algorithm,this paper studies instance feature pooling and loss function on the basis of qualitative and quantitative analysis of MI-Net and mi-Net algorithm.By improving MI-Net and mi-Net algorithm structure and bag loss sorting function,the performance of Deep Multiple Instance classification algorithm is improved,and the following innovative achievements have been achieved.(1)A Deep Multiple Instance Learning method with instance score as weight is proposed.The method based on the embedded space more MI-Net and instance neural network based on the instance space mi-Net instance neural network algorithm on the basis of the different weights of instance distribution inside the bag,and the instance points with the instance weight pooling get bag,the bag represent says give full play to the weight of contribution,can undertake adjustment according to the task and the data,Thus,better results can be obtained and verified by experiments.(2)A bag classification network based on pairwise bag loss is proposed.In this network,pairwise bag loss function is introduced based on Bag Similarity Instance Learning Network architecture based on bag space.Then,the model based on AUC evaluation index is modeled.On the basis of reserving and selecting a pair of positive and negative bag for iteration,positive and negative inter-bag relations are introduced into the loss function to improve the utilization rate of bag.Finally,the experimental results show that the proposed method can improve the sample utilization rate,achieve better convergence effect for network,and improve the accuracy of bag classification.