Research On Scarce Image Classification Algorithms Based On Deep Learning

Deep learning techniques have a wide range of applications in the field of image classification,and usually deep learning algorithms need a large number of labeled samples for training.However,in some application scenarios,there are fewer or even no samples.As a result,the image classification in this sample scarcity situation is poor.Therefore,scarce image classification becomes an urgent problem to be solved.Scarce image classification,also known as image classification in lowresource scenarios,mainly contains few-shot,zero-shot,and mixed scenarios,and has a wide range of applications in image content security auditing area.In this thesis,scarce image classification algorithms in different scenarios are investigated,and the main work includes:1.The thesis constructs a scarce image classification dataset based on actual security and stability business scenarios,which contains a series of subcategories related to security and stability and can be used for the training of classification models for actual business scenarios.The thesis conducts parameter fine-tuning training based on currently popular pretrained models to obtain benchmark classification results of each pretraining model under this dataset,which facilitates the comparison of experimental results with several improved algorithms proposed subsequently in the thesis.2.For the scarce image classification task in the few-shot scenario,a graph neural network model based on distribution self-aggreation is proposed in order to solve the problem of inadequate utilization of sample feature distribution information.The distribution self-aggreation module can automatically learn the similarity among distribution features,and it is experimentally demonstrated that the module can enable effective propagation of sample features,thus improving the classification of scarce images in few-shot scenario.3.For the scarce image classification task in the zero-sample scenario,in order to solve the problems of insufficient ability of existing models to extract prior knowledge and unreasonable cross-modal mapping,the thesis proposes a graph convolutional network model based on multi-headed attention mechanism and feature transformation mechanism.It is demonstrated that the multi-head attention mechanism can effectively obtain the information of graph nodes and the feature transformation mechanism can optimize the cross-modal mapping method,thus improving the classification effect of scarce images in zero-sample scenario.4.For the scarce image classification task in the mixed scenario,a scarce image classification model based on a large-scale multimodal pretrained model is proposed in order to address the problem of inadequate utilization of prior knowledge by existing models.The model introduces image-text description pairs as prior knowledge and incorporates a learnable convolutional network to fine-tune the model.Experiments demonstrate that the model is generalizable for scarce image classification and can solve both zero-shot and few-shot problems..The thesis proposes various innovative algorithms based on deep learning techniques for different scenarios of scarce image classification.The effectiveness of the algorithms is demonstrated through extensive experiments and analyzed on the actual collected scarce image dataset,which is also deployed to a practical application system.Therefore,the research of the thesis has certain theoretical significance and application value.