Construction method of radar target detection model

Abstract

The invention provides a method for constructing a radar target detection model. The method comprises the following steps: acquiring original radar wave data; clutter suppression is carried out; performing imaging operation to obtain a two-dimensional distance Doppler map; performing feature enhancement and multi-scale feature fusion to complete target extraction; recording a size parameter of the target bounding box, calculating a loss parameter and a judgment score of the network model, returning to carry out iterative updating when the judgment score is greater than a preset threshold value, and otherwise, outputting a target detection model; wherein a local response normalization layer of the GoogLeNet network is replaced by a batch normalization layer during iterative updating. According to the method, clutter suppression is carried out on training data, then feature extraction is carried out based on the GoogLeNet network to realize extraction of a target frame, a local response normalization layer is replaced by a batch normalization layer, the convergence speed of a network model is accelerated, and a small sample training over-fitting phenomenon is suppressed, so that under small sample training, the robustness of the target frame is improved. And the detection model meeting the requirements can be quickly obtained.

Description

technical field [0001] The invention relates to the technical field of radar target detection, in particular to a method for constructing a radar target golden policy model. Background technique [0002] Traditional target detection algorithms are highly dependent on artificially designed features, and mainly use intuitive sense or heuristic prior knowledge such as chromaticity contrast, background comparison, and boundary point priors to detect targets. The processing features are relatively single, and it is difficult to deal with multiple targets in complex electromagnetic environments. Detection is time-consuming and labor-intensive, resulting in poor applicability. [0003] At present, there are many researches on target detection methods based on deep learning at home and abroad, and the learning algorithms are rich, but there are still several important problems that have not been solved. The generalization performance of the network model is not strong, and it is di...

AI Technical Summary

Problems solved by technology

[0002] Traditional target detection algorithms rely heavily on artificially designed features. They mainly use intuitive perception or heuristic prior knowledge such as chromaticity contrast, background comparison, and boundary point prior knowledge to detect targets. The processing features are relatively single, and it is difficult to deal with multiple targets in complex electromagnetic environments. Detection, time-consuming and labor-intensive, resulting in poor applicability

[0003] At present, there are many researches on target detection methods based on deep learning at home and abroad, and the learning algorithms are rich, but there are still several important problems that have not been resolved. ability, the generalization performance of the network model is not strong, it is difficult to adapt to target detection in different complex scenarios

Second, the existing deep learning-based region detection algorithm has a large amount of training sample data, complex network algorithms, large calculations and too many candidate boxes, resulting in low detection efficiency, which is difficult to meet the requirements of practical engineering applications

[0004] The invention patent application with the publication number CN113486961A discloses a deep learning-based radar RD image target detection method with a low signal-to-noise ratio. The RD image is manually marked based on the radar wave, and the deep learning network is used for learning to obtain the detection neural network. Realize the detection of the target, but this method still cannot meet the target detection requirements in complex scenes, and the amount of data required for training is large, which is difficult to meet the requirements of actual use

Images