Research On Optimization Method Of Real-time Detection Network For Targets On Water Surface

Nowadays,with the rapid development of artificial intelligence technology,unmanned surface vehicles(USVs)have attracted an increasing world-wide attention for its role in intelligent marine equipment.The key technology of real-time surface target detection is of great significance in promoting the efficiency of USVs’ performance,such as navigation obstacle avoidance,environmental survey,and maritime search and rescue.Unfortunately,there are many severe conditions in the field of USVs,such as variable surface environment conditions,multiple detection targets,severe turbulence when sailing,detecting small targets on the high-resolution images and the limited performance of embedded computing platforms,etc.These problems make it difficult for the speed and accuracy of surface target detection to meet actual application requirements.According to the features of surface target images and the demand of real-time detection on embedded computing platform,a real-time surface target detection network based on deep learning has been proposed.The detection accuracy and speed of the network are improved through network optimization methods such as network structure optimization,multi-scale training,network pruning,knowledge distillation,etc.The outcome has been tested on the USVs.This thesis focuses on the surface targets from the perspective of USVs and the realtime surface target detection method.In order to improve the network’s environmental immunity to interference,a surface target data set has been established using pictures taken at the water surface during various time periods,weather and noises.The detection accuracy and speed of the network are improved by adding pyramidal pooling layers to the network structure,improving the Io U(Intersection-over-Union)formula in the loss function,freeze training and mixed precision training.Faced with the problem that small target detection is difficult under high-resolution images,the network structure is improved by increasing the number of output layers and decreasing the number of down-samples.In addition,the network is trained on multiple scales to improve the accuracy of small target detection.Based on the network model compression technique,the BN(Batch Normalization)layer parameter is introduced as the L1 regularization term in the loss function to realize the network thinning.The channel pruning and Res module(residual module in Res Net)pruning strategies are used to prune the sparse network,which realizes the network model compression and further improves the detection speed of the network.Finally,the pruning network is fine-tuned by knowledge distillation to restore the detection accuracy of the network.The YOLO-Ship network in this thesis shows excellent detection results in the actual test.Compared with the YOLOv3 algorithm,the m AP(mean Average Precision)is improved from 54 to 71,the detection speed is improved by 2.67 times,the network model size is only 7.25% of YOLOv3,and the operating frame rate on the Jetson AGX Xavier reaches 20 fps.The network detection performance exceeds other detection networks,which provides a practical reference for solving the problem of real-time target detection on USVs.