Lightweight Object Detection On The Sea Surface Based On Unmanned Surface Vehicle

Object detection on the sea surface plays a key role in the development and utilization of marine resources.The sea environment is complex and changeable,and there are many kinds of objects.Considering the factors such as safety and obstacle avoidance of unmanned surface vehicle,the shooting process of sea surface object detection images will make the number of small and medium-sized objects in the image majority,which puts forward higher requirements for accurate detection of objects on the sea surface.At present,the object detection method based on deep learning pays more attention to the construction of deeper network to improve detection accuracy.The network model usually has the problems of too large parameters and complex calculation,which lead to the slow detection speed.In addition,the deep network requires the equipment to have higher computing capacity,and it will affect the detection accuracy of the network if the model is compressed.It is an urgent problem to be solved that the model can ensure or even improve the detection accuracy while being lightweight.In this dissertation,Unmanned Surface Vehicle(USV)is used as the carrier of object detection on the sea surface.For the requirements of obstacle avoidance,intelligent ship and other applications,this dissertation studies object detection on the sea surface by using the strong advantages of deep learning in the field of object detection.The main works are summarized as follows.(1)By comparing and analyzing the detection accuracy and model size of different feature extraction networks,the depthwise separable convolution feature extraction network based on Xception was improved by using the method of model pruning.It was also applied to Single Shot Multi Box Detector(SSD)model,and trained by the idea of transfer learning,a lightweight sea surface object detection model XSSD was obtained.(2)In order to solve the problem of detection accuracy loss and small target miss caused by the reduction of parameters after lightweight,a lightweight attention mechanism module LAM was introduced by analyzing the attention mechanism in computer vision,and applied to feature extraction network of XSSD.Therefore,the improved lightweight sea surface object detection model LAM-XSSD with attention mechanism,was built.The effect of the lightweight attention mechanism module was verified by ablation experiment,and several groups of experiments were carried out with other models respectively,which verified the reliability of LAM-XSSD model.(3)This dissertation reconstructed the Marine Obstacle Detection Dataset in the literature,sets the detection objects as boat and buoy,uses label Img software for calibration,and processes it into VOC2007 dataset format.It is to avoid the over-fitting problem caused by the strong continuity of the original image after frame processing.(4)A sea surface object detection system is designed and implemented.Use axios,Vue.js,element UI and other function libraries and frameworks to complete the browser client page design.Integrate front and back development based on Flask framework and B/S mode.No installation is required,remote access is supported,web-based interactive application services can be provided for users,and API-based online open application services can be provided for front-end and back-end developers,which improves the efficiency of later maintenance and management.