Few-sample data amplification method for target detection

Abstract

The invention relates to a few-sample data amplification method for target detection. The few-sample data amplification method comprises the following steps: S1, qualitatively analyzing possible scenes of a target sample and styles of the scenes; S2, extracting scene features and style features of the target sample by using a scene detection algorithm and a style extraction algorithm; S3, performing clustering analysis on the scene features and the style features by using a machine learning algorithm; S4, searching pictures with similar scenes or styles from the open source data set accordingto a clustering analysis result; and S5, fusing the target sample with the similar pictures found from the open source data set to manufacture a false sample. According to the method, the scene (background) of the target sample can be changed, more false samples with similar styles can be generated so as to make up for the imbalance problem of part of samples in the aspect of sample number, and meanwhile, the problem that training is needed based on a GAN network can be avoided. According to the method, the data amplification efficiency is improved, and meanwhile, the data quality of the sample is also improved.

Description

technical field [0001] The invention relates to the technical field of target detection, in particular to a small-sample data amplification method for target detection. Background technique [0002] Data augmentation is also called data augmentation, which means to make limited data generate value equivalent to more data without substantially increasing data. It is a commonly used data processing method in the field of target detection. Data augmentation plays a very important role in improving model accuracy and improving model generalization ability. Generally speaking, data enhancement can be divided into supervised data enhancement and unsupervised data enhancement methods. [0003] 1. Supervised data enhancement, that is, using preset data transformation rules to augment data on the basis of existing data, including geometric operations and color transformations. [0004] (1) The geometric transformation class is to perform geometric transformation on the image, inclu...

AI Technical Summary

Problems solved by technology

[0005] (2) The geometric transformation operation above the color transformation class does not change the content of the image itself, it may select a part of the image or redistribute the pixels

[0009] In the current field of deep learning, supervised data augmentation is widely used, such as rotation, cropping, deformation, scaling, adding noise, blurring, and color transformation, but these data augmentation techniques cannot change the foreground and background of the target sample. Relative structure, but also can not change the relationship between different target sample size

However, simply increasing the data samples of a few-sample target, such as copying, cannot change the diversity of the target sample data. At the same time, a large number of invalid samples are added, which increases a lot of learning time in vain.

[0010] The use of unsupervised data enhancement methods such as GAN networks can greatly increase the data of few samples and increase the richness of samples, but the disadvantages are also very obvious: a pair of generative confrontation networks or other learning models need to be trained, and generative confrontation The network or other learning models themselves require a large number of samples for training, which directly increases the difficulty of the algorithm and makes it difficult to implement

Images