Dynamic label smooth weighted loss method and device based on historical records

Abstract

The invention provides a dynamic label smooth weighted loss method and device based on historical records, and belongs to the technical field of image processing. The method comprises the following steps: S1, building a historical queue with a fixed length for each image, and initial values in the historical queues are real label categories of the corresponding images; s2, during each iterative training of the machine learning model, calculating weights of corresponding images belonging to a real category and other categories according to a historical queue, and combining the obtained weightswith a cross entropy function to calculate loss so as to optimize the machine learning model; and S3, after each iterative training is completed, updating the corresponding historical queue by using the category to which each image belongs predicted by the machine learning model, and returning to S2 to continue iteration until the training is completed. By adopting the invention, the problem of low recognition accuracy of the machine learning model due to similarity among classes in the classification task can be solved.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a method and device for dynamic label smoothing and weighted loss based on historical records. Background technique [0002] Classification is a commonly used form of data analysis in data mining and is usually applied to predicting categories (or probabilities of categories) of data rather than continuous numerical values. In practical applications, limited by the understanding of the data or the requirements of the task, there are similarities between different types of data, which affects the learning process of the data mining model and causes the model to output wrong classification results. [0003] Taking the application scenario of cell cancer image data classification as an example, cell cancer is a continuous process, and it is impossible to set an absolute threshold to quantitatively distinguish which stage the current cell belongs to. For example, th...

AI Technical Summary

Problems solved by technology

[0005] Embodiments of the present invention provide a dynamic label smoothing weighted loss method and device based on historical records, which can solve the problem of low recognition accuracy of machine learning models caused by similarities between categories in classification tasks. The technical solution is as follows:

Images