Research Of Few-shot Dermatoscope Image Classification Based On Metric Learning

The classification of skin diseases from images is the key to the diagnosis of dermatology.The development of computer-aided detection system can effectively reduce the pressure of doctors and improve the efficiency of diagnosis.Deep learning has made great achievements in the field of computer vision.However,excellent deep learning models are based on large-scale and high-quality labeled data.For privacy and moral considerations,it is very difficult to obtain large-scale labeled samples in the medical field.When the training samples are insufficient,the model will have problems such as overfitting and poor generalization ability.Few-shot learning has the ability to learn quickly,can learn rich target features from a small amount of label information,and complete new sample-related tasks based on past experience.Metric learning is an effective method to solve few-shot learning problems.However,due to the long tail phenomenon of skin disease data set distribution and the high similarity between categories,few-shot learning is not effective in classification tasks.This paper proposes an improved method for few-shot skin disease image classification.The main contents include the following points:(1)Decompose the metric learning process and establish a three-stage learning paradigm.In the first stage,the embedded module is used for feature extraction.In the second stage,the metric space is established.In the third stage,the measure module is used for measuring the distance between the query set sample and each type of samples.(2)The enhanced embedded module is obtained by two-stage training.The first stage training uses the whole training set to enhance the feature extraction ability of the embedded module.The second stage is training in the form of meta-tasks for enhancing the generalization ability of the embedded module in the test stage.Finally,the selfdistillation mechanism is used to improve the overall performance of the model,and the optimal iterative model is got through experiments.(3)The subspace method is introduced to establish the metric space,and the category features are well summarized,which effectively reduces the search space for distance measurement of query set samples,improves the classification accuracy,in the meantime,improves the training efficiency of the network,and reduces the calculation of parameters.(4)Construct a multi-scale distance measure module,which combines the advantages of Euclidean distance and Cosine distance.In the calculation of distance,both the direction and the absolute value of distance are considered.It can be effectively calculated in low-dimensional and high-dimensional spaces to improve the accuracy of classification.(5)Using generative adversarial networks to alleviate the problem of extreme imbalance of data sets,expand the small sample categories distributed at the tail.The autoencoder with embedded module is used to improve the generative adversarial network.The information of the autoencoder is embedded into the generator and the discriminator,and the loss function of the discriminator is improved by gradient penalty based on prior knowledge.