Research Of Few-Shot Image Classification Based On Saliency Enhancement And Bias Correction

The success of artificial intelligence has brought great convenience to people.Image classification as one of the important tasks also gets great attention.But the training is often inseparable from the massive labeled data.However,annotating adequate data is expensive and time-consuming,and there are many scenarios such as medical images that are difficult to obtain sufficient training data.Therefore,it’s promising to develop image classification in the condition of small data.This dissertation takes the few-shot image classification as the main research target.Task-based training pipeline is the main stream in current methods.Multiple few-shot classification tasks which are randomly sampled from the training set with the same parameters as the test phase are used to train the model.This kind of framework often leads to two problems:(1)Due to the small amount of data in each task,the outliers which are selected during the random process will have great negative impact on the accuracy.(2)The essence of task-by-task training is carried out on a fixed data set,and there is still a strong bias on the training set.To tackle the above problems,this dissertation studies the few-shot classification based on the saliency enhancement and bias correction which can correct the feature distribution including outliers and the training bias.The contents and contributions are as follows:This dissertation corrects the outliers by comprehensively considering the intraclass similarity and inter-class uniqueness of samples.Firstly,a position-weighted attention module is proposed to measure the discriminativeness of different regions,so as to generate the important local feature for each image.It can describe the training data more completely.Secondly,an instance-weighted attention module is applied to calculate the intra-and inter-class attention values of each embedding,and further assigns smaller weights to the low-quality images.Finally,an adaptive prototype is computed for each category to enhance the saliency of features.Self-supervision is a common idea to correct the model’s bias for the training set and improve the generalization.Based on the idea that the output between the perturbed data should be consistent,A multi-level auxiliary loss is introduced in this dissertation.Given the sampled classification task,the model can obtain two copies of the same task with the help of different data augmentations.From the sample level,the relationships between classes of each image from two copies are forced to be consistent.From the task level,the prototypes from the two copies are constrained to be the same.They act as the second task to assist the training of the model and to mine the inherent characteristics of the data.This dissertation proposes an attention-based feature saliency enhancement network and a self-supervision based multi-level consistency loss which can identify the outliers and correct the bias effectively to improve the few-shot image classification task.