Data enhancement method for small-scale target

Abstract

The invention provides the data enhancement method for the small-scale target, the quality of the inserted image is improved by using a copying and pasting mode and a triple linear interpolation method, the problem of sample imbalance of the small-scale target in a neural network is solved, and the accuracy of the data enhancement method is improved. The function of improving the model performance of small-scale target detection and semantic segmentation based on the deep learning technology is realized. On the premise of not changing basic structures (such as textures, objects, context semantic environments and the like) of images, the proportion of small-scale targets in the training data set is effectively increased, the weight of the small-scale data set in deep neural network training is improved, and the problem of sample imbalance is alleviated. The method is used for establishment and enhancement of data sets in actual scenes such as focus detection in clinical medicine, finding and tracking of suspicious targets in the air in military and the like, the overall performance of detection and segmentation of small-scale targets is improved, later model training is helped to obtain better scores, and the robustness of the model is enhanced.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to a data enhancement method for small-scale targets. Background technique [0002] Object detection is an important research direction in the field of computer vision and the basis of other complex vision tasks. In recent years, with the rapid development of deep learning (network structures such as Alexnet, ENet, UNet, FastSCNN, etc.) and the use of some data enhancement methods, target detection has achieved great success in both accuracy and speed. [0003] However, compared with regular-sized targets, small-scale target detection has long been a difficult point in target detection. Problems in small-scale target detection include few available features, small proportion in existing data sets, and unbalanced samples. This has also led to the fact that on some public detection datasets such as MS COCO, the performance of small-scale target detection is usually ...

AI Technical Summary

Problems solved by technology

[0003] However, compared with conventional-sized targets, small-scale target detection has long been a difficult point in target detection. The problems of small-scale target detection include few available features, small proportion in existing data sets, and unbalanced samples.

This also leads to the fact that on some public detection datasets such as MS COCO, the performance of small-scale target detection is usually less than half of that of large targets.

The current popular data enhancement methods are to rotate, scale, color transform, flip, etc. on the image. However, these data enhancement methods start from the perspective of the whole picture, not from the target level, which leads to these methods in small-scale targets. Poor performance of separation precision

[0004] Traditional data enhancement methods, such as rotation, scaling and flipping, etc., these methods will not increase the number of small-scale samples. The defects of traditional data enhancement methods make them unable to solve the problem that the number of small-scale samples is too small in the entire sample. In order to be more Accurate segmentation of small-scale targets requires an algorithm to solve the problem of small-scale targets in the data set and unbalanced samples

Images