Sample data acquisition method, acquisition system, server and computer readable medium

Abstract

The present disclosure provides a sample data acquisition method, including: constructing a mother sample picture database; sampling the mother sample picture database multiple times to obtain multiple corresponding mother sample picture sets; for each mother sample picture set, using A marquee of a predetermined size extracts a plurality of sub-sample pictures from each mother sample picture in the mother sample picture set, and assigns a preliminary class label to each sub-sample picture, so as to obtain each master sample picture set. Corresponding sub-sample picture set; for each sub-sample picture set, use all the sub-sample pictures contained in the sub-sample picture set as training sample data, and train the sample classification model corresponding to each sub-sample picture set; for each sub-sample picture set A sub-sample picture, the sub-sample picture is respectively input into each sample classification model, and the classification result with the largest frequency is selected as the calibration class mark of the sub-sample picture.

Description

technical field [0001] The present disclosure relates to the field of deep learning, and in particular, to a sample data acquisition method, acquisition system, server, and computer-readable medium. Background technique [0002] When training a detection model for a specific task based on deep learning (Deep Learning) technology, it is necessary to collect a large number of training sample data with calibrated categories in advance. [0003] However, in practical applications, it is found that for some special tasks, it is difficult to obtain a large number of labeled (with labeled categories) small-sized samples; for example, in the lesion detection task for fundus images, the trained detection model can To detect whether there is a lesion in the fundus image, and to locate the lesion when the lesion is detected, it is necessary to obtain a large number of small-size (patch)-level or pixel-level lesion-level sample data. Currently, it can only be done manually. Sample coll...

AI Technical Summary

Problems solved by technology

[0003] However, in practical applications, it is found that for some special tasks, it is difficult to obtain a large number of labeled (with labeled categories) small-sized samples; for example, in the lesion detection task for fundus images, the trained detection model can To detect whether there is a lesion in the fundus image, and to locate the lesion when the lesion is detected, it is necessary to obtain a large number of small-size (patch)-level or pixel-level lesion-level sample data. Currently, it can only be done manually. Sample collection by selecting and marking in fundus images

The above artificial sampling method has the following problems: 1) Due to the complex shape of the lesion, the doctor's labeling is highly subjective, and the boundary division is relatively random. It is also difficult for professional ophthalmologists to define the attributes of the lesion boundary pixels, that is, the labeling is difficult; 2. ) It takes a long time for doctors to manually label, and the cost of obtaining the calibration category of samples is high, that is, it is difficult to obtain a large number of samples

Images