Attribute-Based Classification Algorithm For Few-Shot And Application To Medical Images

Deep learning has made remarkable achievements in image recognition.However,traditional deep learning methods require a considerable amount of labeled data,which can be a time-consuming and labor-intensive process for annotation.The lack of datasets in traditional deep learning tends to bring about problems of overfitting and poor generalisation.To address this class of problems,a method for obtaining higher performance and generalisation of deep learning models with small amounts of data is called small sample learning.The emergence of small sample learning has brought a revolutionary solution to these problems,small sample learning means.In this paper,we conduct algorithmic research on few-shot learning from image attributes and work on applying few-shot learning methods to medical imaging.The main work in this paper is as follows.(1)To address the issue of inaccurate classification resulting from the insufficient amount of traditional metric learning data,this paper introduces an algorithm,AMCNet,which leverages image attribute information to enhance the accuracy of metric classification,i.e.to establish a new classification similarity using attributes based on the original overall prototype feature similarity of traditional small samples.The model acquires image attributes from the training set and then uses the acquired attributes to establish a link between the queries in the new class and the test set samples;this new link is defined as the attribute distribution similarity.In this paper,we use attribute distribution similarity to complement overall feature similarity to compensate for the inaccurate classification of overall feature similarity in some cases,and obtain more accurate classification results.Through experiments,it has been demonstrated that the proposed method in this paper outperforms most of the other related few-shot classification algorithms in terms of accuracy on the mini Image Net,and CUB datasets.(2)To address issues such as insufficient medical image data and poor model generalization ability,we plan to adopt the method of few-learning to alleviate such situations.The text introduces a self-supervised pre-training method based on contrast learning from the perspective of image foreground and hindground attributes of medical datasets,and adds a new module to this basic method to propose a medical few-shot classification algorithm based on attribute contrast learning.The algorithm differs from previous supervised pre-training methods widely used on standard datasets in that the self-supervised pre-training under contrast learning used in this algorithm performs more prominently.In addition,the module proposed in this paper enables the generalisation capability and performance of the model to be improved.It is shown through experiments that the proposed method performs better on the ISIC2018 and SPPH datasets compared to the current mainstream methods on small samples in medicine.