RD time-frequency data-oriented deep learning model evaluation system and method

Abstract

The invention discloses a deep learning model evaluation system and method oriented to RD time-frequency data. The process comprises verification set expansion, multi-IOU threshold value F1 calculation, multi-IOU threshold value mAP calculation, FLOPs calculation and evaluation index integration calculation. The main body of the evaluation method is a verification set expansion method and an integration strategy-based model evaluation method, a verification set is expanded through an image fusion and inspection mechanism, and the problem of RD-oriented time-frequency data shortage is solved; a larger-scale verification set can be obtained under the conditions that obvious noise information is not introduced, target label information is not leaked and data distribution is not changed, so that the verification set better represents the characteristics of overall data, and the generalization performance of the model is better evaluated; the overall capability of the deep learning model is represented by the F1 score of the multiple IOU threshold values, the mAP of the multiple IOU threshold values and the weighted score of the FLOPs, and the capabilities of the model in the aspects of single-point optimization, global average optimization and time performance can be represented more accurately, so a powerful technical support is provided for the evaluation of the deep learning model facing RD time-frequency data.

Description

technical field [0001] The invention relates to the fields of aerospace and artificial intelligence, in particular to a deep learning model evaluation system and method for RD time-frequency data. Background technique [0002] With the continuous application and rapid development of deep learning in computer vision such as image target detection, deep learning has defeated traditional methods in more and more fields. Under the premise of a small amount, the use of deep learning can achieve excellent results. RD time-frequency data is collected by radar and other devices, and images with physical meaning can be obtained through time-frequency analysis and image rendering. In the application of RD time-frequency data, object detection is one of the most concerned areas. Traditional target detection methods based on RD time-frequency data often involve more analysis and judgment based on experience, and corresponding analysis methods need to be adopted according to specific d...

AI Technical Summary

Problems solved by technology

[0004] At present, there are many image augmentation methods at home and abroad, but there is no verification set expansion method for phase-oriented RD time-frequency data images. Domestic and foreign methods only use F1 score under multiple IOU thresholds, mAP and FLOPs under multiple IOU thresholds. One is relatively common as a model evaluation index, but there is no example of using weighted scores to represent the overall effect of the model

Images