Complex battlefield environment target efficient identification method based on improved FasterR-CNN

Abstract

The invention discloses a complex battlefield environment target efficient identification method based on an improved FasterR-CNN. The method comprises the following steps: constructing two paths of feature extraction networks; inputting a battlefield environment feature map output by one path of feature extraction network into a candidate region generation network, and enabling the candidate region generation network to distinguish a background and a target in a battlefield environment by fusing shallow target position information and deep high-order semantic features; resetting the anchor point size and the length-width ratio of the battlefield environment feature map by adopting a K-Means clustering method; and inputting the battlefield environment feature map output by the other path of feature extraction network and the candidate region output by the candidate region generation network into a classification regression network, performing classification identification on the target, and performing regression on the position. According to the method, in a complex battlefield environment, the accuracy of a battlefield target is improved, so that the model has better theoretical guidance significance compared with a previous model.

Description

technical field [0001] The invention belongs to the field of deep learning and image recognition, in particular to an efficient recognition method for complex battlefield environment targets based on improved FasterR-CNN. Background technique [0002] In a complex battlefield environment, how to effectively and accurately identify battlefield targets for unmanned equipment clusters is the key to achieving precise control of unmanned equipment, obstacle avoidance, and execution of attack tasks. The identification of enemy targets can ensure that our unmanned Complete reconnaissance and combat missions off the battlefield. Traditional object recognition methods mainly include Cascade+HOG, DPM+Haar, SVM and their improvement and optimization methods. The disadvantage of these methods is that they need to manually design features, and the workload is huge, and similar algorithms have large changes in target shape, complex background or light Insufficient cases tend to perform p...

AI Technical Summary

Problems solved by technology

The area generation network needs to generate candidate areas. In the original convolutional network, the feature image output by the last convolutional layer of the ZF or VGG network is directly used to predict the foreground position. Due to the low resolution of the feature image, it is serious for small objects or being occluded. The characteristics of the target will be lost

Images