A Deep Compression Neural Network Algorithm for Unmanned Surface Vehicle Target Recognition

For the target recognition requirement of unmanned surface vehicles (USV) in complex environments, a recognition algorithm based on deep compression neural network is proposed. The proposed algorithm employs VGG-based network to extract features, and improves the bounding boxes matching strategy and loss function of SSD detection algorithm. The clustering algorithm is used to optimize the recognition process and improves the recognition accuracy. The multiple feature maps are combined to achieve robust recognition of multi-scale target quickly. Finally, the memory storages of the proposed algorithm are greatly reduced, because the network is compressed by 50% without affecting the performance by using the deep compression method. The algorithm is implemented and verified on the embedded GPU NVIDIA Jetson TX2 platform. The experiment results show that the algorithm can recognition multiple classes of targets in real-time and multi-scale in complex environments. Besides, the proposed method has strong robustness to weather and illumination changes and the recognition speed of single-frame video is 0.1 second. © 2019, Editorial Department of Transaction of Beijing Institute of Technology. All right reserved.