| Image classification task,as an important topic in the field of computer vision,aims to predict the semantic category of a given image.Accurate image semantics are the foundation of downstream tasks such as object detection,semantic segmentation,and image generation.How to improve image classification accuracy in situations where data features are difficult to learn,data annotation quality is low,and the number of data annotations is small is a key issue to be solved in image classification tasks.There are two types of methods for image classification: deep learning and non deep learning.Deep learning methods exhibit excellent classification performance based on large-scale annotated data,but large-scale annotated datasets require significant human and material resources,and labeling a large number of samples for each category does not conform to the objective laws of objects in the real world.Therefore,how to make correct predictions under limited data volume and limited annotation is one of the important issues to be solved in image classification tasks.Dictionary learning is a traditional and high-precision image classification method with a small number of model parameters and a low demand for annotated sample size.This paper focuses on dictionary learning as the research core,with the goal of high-precision image classification,and mainly completes the following aspects of work:Firstly,in response to the issue of traditional dictionary learning methods treating each training sample equally,resulting in redundant information extraction from simple samples and insufficient information extraction from complex samples,this paper proposes an adaptive enhanced dictionary learning method.This method trains multiple dictionaries and weak classifiers that differentially focus on samples,adjusts the weight of the samples based on the classification results of the model,and achieves the goal of enhancing the model’s attention to difficult samples.The experimental results show that this method achieves high-precision image classification on facial image classification,object classification,and scene classification datasets by integrating multiple dictionaries and weak classifiers.Secondly,to address the issue of classification methods that focus on difficult samples being susceptible to label noise interference,this paper proposes a deep residual discrimination dictionary learning method.This method adopts a divide and conquer approach to learn multi-layer discriminant dictionaries.The learning goal of high-level dictionaries is to fit the residuals of low-level dictionaries.The multi-layer residuals learning method ensures that low-level models that are less affected by noise labels always dominate category prediction,thereby achieving attention to difficult samples while avoiding model learning bias caused by noise label samples.Experiments have shown that compared to other classification methods that focus on difficult samples,this method achieves high-precision image classification and exhibits stronger robustness in the presence of noisy labeled samples.Thirdly,in response to traditional cross domain dictionary learning methods that typically establish visual semantic embedding relationships at the class level while ignoring fine-grained information at the image level,this paper proposes a zero-shot image classification method based on hierarchical semantic alignment cross domain dictionary learning.This method first learns the visual category center of each image,and then uses the visual category center and category semantic description of the image as input to learn cross domain dictionary pairs,so that the visual space and semantic space are projected into the same embedding space through their respective dictionaries.The semantic vector of the known category image is generated by using the known category image and the cross domain dictionary pair obtained from the training.Finally,the cross domain dictionary pair of image layer is trained by using the visual features of known category images and their corresponding semantic vectors.The experimental results show that this method achieves high-precision prediction on zero-shot learning tasks through hierarchical visual semantic embedding from coarse to fine.Fourthly,in response to existing methods for zero-shot learning tasks that establish an embedding relationship between visual features and category semantic vectors,while neglecting fine-grained semantic knowledge in semantic vectors,this paper proposes a zero-shot image classification method based on common feature perception cross domain dictionary learning.This method introduces attribute features to mine fine-grained semantic knowledge of commonalities between different categories.Firstly,visual features of commonalities between categories are extracted based on attribute features,and then the degree to which an attribute is included in the image is simulated through expert scoring to form the semantic vector of the image.Finally,a fine-grained visual semantic cross domain dictionary is constructed using image visual features and image semantic vectors to effectively transfer knowledge from known categories to unknown categories.The experimental results show that this method has higher classification accuracy compared to other zero-shot learning methods.In summary,this article aims to conduct research on multi dictionary learning with the goal of high-precision image classification.Various multiple dictionary learning models have been proposed for different feature domains of test data and training data,achieving more user-friendly and user-friendly image classification,and expanding its application scope to a certain extent. |
