A method and application of image sample generation for target detection

Abstract

The invention discloses an image sample generation method and application for target detection. The method includes: partitioning the foreground sample and / or the background sample to obtain a plurality of image sample mosaic areas, and respectively establishing a plurality of image sample mosaic areas Limit the out-of-bounds area; use the image data enhancement strategy search space to enhance the data of positive samples and negative samples; paste the enhanced positive samples and negative samples into different image sample splicing areas of foreground samples and background samples, according to the positive The position of the samples and negative samples in the foreground samples and background samples, calculate the coordinates of the target detection frame, obtain the enhanced foreground samples, background samples, and the data labels of the enhanced foreground samples and background samples; Mixed samples are data augmented using crippled augmentation strategies. The present invention is suitable for enhancing data in a scene with a small amount of data, realizing the training of a neural network, and improving the MAP and AP50 of the model.

Description

technical field [0001] The present invention relates to the technical field of target detection, more specifically, an image sample generation method and application for target detection. Background technique [0002] With the vigorous development of artificial intelligence technology, deep learning technology has made breakthroughs in the fields of classification, recognition, detection, tracking, and segmentation in the field of computer vision. Compared with traditional machine vision methods, the deep neural network extracts effective data features from various data sets under the training of a large amount of data rich in local correlation characteristics, which has the characteristics of high precision, strong generalization, and fast speed. Although deep learning is superior to traditional machine learning, deep learning needs to be based on a large amount of relevant data. When training the target network model, if the sample size is insufficient, it will often lead...

AI Technical Summary

Problems solved by technology

[0006] Although this kind of discrimination method is simple and fast, it also has the following defects: the available features of the generated enhanced data are not rich enough, and it is easy to be learned by the neural network; the performance of target classification is improved obviously, but the improvement of target detection is limited

This method can generate data that is related to and different from the original data. The disadvantage is that it is difficult to design a suitable generation method. If the design is not good, the generative adversarial network cannot converge or cannot generate diverse and different from the original image. image, and the training is too complicated, the target and background information in the image is not accurate, that is, the lack of detailed information

Images