Research On The Method Of Robust Partial Labeled Multi-label Image Classification

With the current exponential growth of visual information and the increasing scale and complexity of image data,how to efficiently process and analyze image data has become one of the hot topics in the field of artificial intelligence.At present,image classification has been widely used in many fields,such as medical image diagnosis,security monitoring,intelligent transportation,etc.However,in practical applications,images often have different degrees of noise or bias due to the interference of various factors,either due to ambiguous label annotation or interference of wrong labels,and even slight label perturbations can degrade the performance of traditional deep image multi-label classifiers,and these partial labeled problems make it difficult to achieve the desired results of traditional multi-label image classification methods.Therefore,how to build deep neural network models for partial label and how to solve the impact of label perturbation on multi-label image classification models have become major hot issues to be solved in the field of deep learning.In this thesis,we firstly describe the current status of domestic and international research on multi-label image classification;secondly,we introduce and describe the learning methods of graph convolutional networks and large margin techniques in detail;we take the problem of partial labeled as the research point and start the research on partial labeled multi-label image classification methods,and propose the partial labeled multi-label image classification methods based on multilabel joint feedback graph convolutional networks(GCN)and deep large margin ranking lossbased multi-label image classification methods.These methods enable the model to better deal with the partial label problem and label perturbation problem;finally,based on this,a prototype system for multi-label image classification is designed and implemented.The specific research work in this thesis consists of the following items:(1)Proposed a partial labeled multi-label image classification method based on multi-label joint feedback GCN.Considering that since the label information of the training dataset contains fuzzy labels as well as noisy labels,and the label structure is highly dependent on this information,a partial labeled multi-label image classification method based on multi-label joint feedback GCN is proposed.The method differs from a single statistically constructed adjacency matrix in the construction of the label structure,and adds the prior knowledge of the labels to construct the adjacency matrix to form a GCN-based multi-label learning network to prevent the overfitting phenomenon caused by noisy labels.Meanwhile,a joint feedback mechanism is proposed to overcome the problem of poor classification accuracy caused by biased labels with insufficient information,and the label data is iteratively updated during the training process so that the multilabel classification model can maintain good classification performance even on the labeled sample set containing fuzzy labels as well as noisy labels.Numerous experiments are conducted on relevant multi-label image datasets,and the experimental results verify that the method can effectively reduce the impact of classification accuracy brought by fuzzy and noisy labels,and compare with the frontier methods,the method can effectively improve the comprehensive performance of multi-label image classification.(2)Proposed a partial labeled multi-label image classification method based on depth large margin ranking loss.To address the problem that even the presence of slight label perturbation degrades the performance of traditional deep learning-based image multi-label classifiers,a partial labeled multi-label image classification method based on depth large margin ranking loss is proposed.The method maximizes the distance from each training point to the decision boundary by treating each intermediate layer activation value of the deep learning-based multi-label image classifier as an intermediate representation of the image,using an arbitrary -paradigm to calculate the label margin metric paradigm for the label margin metric,with a view to effectively solving the performance degradation of multi-label classification due to label perturbation.Meanwhile,a large margin loss optimization method based on negative sampling technique is proposed to make the time complexity of the loss function linearly proportional to the size of the label set to adapt to the label-sets expansion scenario.The experimental results show that the proposed method can effectively tolerate label perturbations and maintain high classification performance.Compared with related advanced methods,the classification performance of the proposed method is more robust and can further improve the comprehensive performance of multi-label image classification.(3)Design and implementation of a prototype multi-label image classification system,which uses Py Charm,Python,Py Torch,Open CV,Py Qt’s Qt Core module and Qt Gui module,opencvpython,torchvision and other algorithm libraries and utilizes the proposed methods in(1)and(2),and designs and implements a prototype multi-label image classification system.The system integrates image data preprocessing module,tag graph model building module,model training module and multi-label image classification and visualization module.