Multi-objective recognition based on deep learning

Abstract

Purpose With the continuous development of aerospace technology, space exploration missions have been increasing year by year, and higher requirements have been placed on the upper level rocket. The purpose of this paper is to improve the ability to identify and detect potential targets for upper level rocket. Design/methodology/approach Aiming at the upper-level recognition of space satellites and core components, this paper proposes a deep learning-based spatial multi-target recognition method, which can simultaneously recognize space satellites and core components. First, the implementation framework of spatial multi-target recognition is given. Second, by comparing and analyzing convolutional neural networks, a convolutional neural network model based on YOLOv3 is designed. Finally, seven satellite scale models are constructed based on systems tool kit (STK) and Solidworks. Multi targets, such as nozzle, star sensor, solar,etc., are selected as the recognition objects. Findings By labeling, training and testing the image data set, the accuracy of the proposed method for spatial multi-target recognition is 90.17%, which is improved compared with the recognition accuracy and rate based on the YOLOv1 model, thereby effectively verifying the correctness of the proposed method. Research limitations/implications This paper only recognizes space multi-targets under ideal simulation conditions, but has not fully considered the space multi-target recognition under the more complex space lighting environment, nutation, precession, roll and other motion laws. In the later period, training and detection can be performed by simulating more realistic space lighting environment images or multi-target images taken by upper-level rocket to further verify the feasibility of multi-target recognition algorithms in complex space environments. Practical implications The research in this paper validates that the deep learning-based algorithm to recognize multiple targets in the space environment is feasible in terms of accuracy and rate. Originality/value The paper helps to set up an image data set containing six satellite models in STK and one digital satellite model that simulates spatial illumination changes and spins in Solidworks, and use the characteristics of spatial targets (such as rectangles, circles and lines) to provide prior values to the network convolutional layer.