| The great success of image classification in agriculture,medicine,intelligent transportation,and secure payment cannot be separated from the rapid development of deep learning.However,deep learning relies heavily on large-scale labeled data.Due to insufficient data or the high cost of manual annotation,many real-world scenarios have only limited labeled data,which limits the further development of deep learning in the field of image classification.Few-shot learning methods do not rely on a large number of training samples,thus avoiding the high cost of labeled samples.Few-shot learning uses available information to learn from a limited number of labeled samples and quickly generalizes to new tasks containing a few labeled samples,overcoming the limitation that deep learning requires a large number of labeled data.This thesis focuses on the image classification method based on few-shot learning,mainly focusing on optimization-based methods,metric-based methods,data augmentation-based methods,and practical application scenarios.The details of the research are as follows:1.Few-shot image classification method based on balanced mechanism is proposed to address the huge network parameters,computational effort,and large amount of time consumed by inner and outer loops caused by the deep network in the optimization-based few-shot image classification method.Firstly,simple linear operations are performed on a few original features to obtain more features,and more features are obtained by using fewer parameters to reduce network parameters and computational effort.Secondly,a local cross-channel interaction strategy without dimensionality reduction is used to obtain significant performance gains without increasing the parameters and computational effort,ultimately achieving a balance between performance,network parameters,and computational effort.Finally,an acceleration strategy is used to reduce the time required for new class identification and speed up the adaptation process for few-shot image classification tasks.2.To address the shortcomings of metric-based few-shot image classification methods that use deep networks and random samples to learn metric space,few-shot image classification based on network simplification and teacher knowledge is proposed.The network is simplified twice to reduce network parameters and computation.In the first simplification,the redundancy of the principal component of the feature input is reduced,and the implicit information in the redundant component is extracted instead of directly discarding the redundant components.In the second simplification,the channel and spatial redundancy of the deeper network layer inputs are reduced.Also,teacher knowledge is used to guide the random sampling process to better utilize the limited samples and improve the classification performance.3.To address the problem of complex and costly data generation in data augmentation-based few-shot image classification methods,few-shot image classification method based on graded selection of enhanced pseudo-labels data is proposed by expanding the dataset with unlabeled samples.Firstly,diverse inputs are used to encourage the label network to learn invariant and robust representations.Then,the depthwise over-parameterized convolutional layer and group residual connection with shared parameters are used to accelerate the network training so as to overcome the time-consuming effect of input diversity and finally obtain enhanced pseudo-labels for unlabeled samples.Secondly,the graded pseudo-labeled data selection is proposed to determine the amount of pseudo-labeled data based on the label network’s performance level,which ensures classification accuracy and avoids the huge consumption caused by using all the pseudo-labeled data.4.To address the problem of medical image classification with insufficient data and expensive manual labeling,few-shot medical image classification method based on auxiliary meta-learning is proposed.In the auxiliary training phase,the feature mapping model is jointly trained with external data,and the category prediction probability of misclassification is eliminated,making it easier for the model to learn the distinguishing information and avoiding the degradation of model performance due to the discrepancy of external data.Then,the redundancy of the principal components of some inputs to the feature mapping model is reduced,and the implicit information of the redundant components is extracted.The training process is accelerated using the depthwise over-parameterized convolutional layer and the three-branch structure.In the meta-train phase,the feature mapping model is optimized in the embedding space based on category prototypes and cosine distance.In the meta-test phase,the few labeled samples are used to classify the unknown data.There are 40 figures,39 tables,and 176 references in this dissertation. |
