Design And Implementation Of Automatic Ship Waterline Detection System Based On Deep Learning

In the process of loading and unloading of bulk cargo terminal,the most common method of cargo weight measurement is waterline weight measurement,which is the key basis for cargo handover,settlement,calculation of freight and tax.At present,the detection method of waterline weighing is manual reading,which is subjectively affected by physical health,unit position and other factors,and objectively affected by wind and waves,sunlight,water quality and other environmental impacts.The port inspection department urgently needs a detection device that can quickly and accurately identify the ship’s waterline,which can not only solve the problem of inconvenience,but also solve the problem of accuracy in detection.This paper designs a detection system based on a lightweight deep learning network built to mobile devices.Based on the analysis of different collection schemes,a scheme based on the three-axis mechanical pan-tilt equipped with mobile device collection is proposed,and the semantic segmentation and target detection algorithms based on deep learning are carried to the mobile device all of which will contribute to the realization of the convenience and accuracy of ship waterline detection.In this paper,the deep learning algorithm is applied to the mobile device for the first time to realize the detection of ship waterline.The network adopts a double branch and multi-scale structure,and designs semantic branch and detail branch.DMB residual block is built in semantic branch,which is fused with detail branch through spatial attention and channel attention through multi-scale feature fusion,so that semantic branch guides detail information and detail branch guides semantic information.After testing,the semantic segmentation model achieves an accuracy of 73.64% in the Cityscapes test set and a reasoning speed of 96.9 frames per second on the RTX2080 TI graphics card.Aiming at the datum line and character of ship waterline,an algorithm based on target detection is built,which solves the conversion of the actual pixel size of the image after semantic segmentation,avoids the error caused by manual operation and is more convenient to use.The semantic segmentation and object detection model proposed in this paper is carried to the mobile device,which solves the problems of inconvenient collection and accuracy of ship waterline detection.Besides,a semantic segmentation dataset and a character detection dataset are created separately for model training.In this study,10 videos of different ships are collected in different environments.Each video produces 200 to 500 semantic segmentation datasets and 100 character detection datasets.The segmented datasets contain a total of 3,000 labels and 1000 character detection datasets.The accuracy of the semantic segmentation model network has reached 99.3% by testing.The rationality of the scheme and the accuracy of the test results has been verified by field testing in Qinhuangdao Port.