Ship Waterline Detection Algorithm Based On Convolutional Neural Network

In the bulk cargo transaction in the port,the ship’s eating line is a certificate for direct weighing transactions.The accuracy of the water line reading is directly related to the fairness of the transaction.At present,port ship line detection mainly relies on manual visual inspection.Such a method is easy to cause disputes due to human factors;at the same time,manual measurement is slow and cannot work when the waves are large.Therefore,how to minimize the artificial and environmental factors to interfere with the detection of the waterline,and how to quickly detect the water gauge line has been an urgent need for the port department to solve two problems.In this paper,a set of ship water gauge video acquisition device is designed,and the advantages and disadvantages of the acquisition scheme adopted in different stages are analyzed.The final adopted magnetic crawler wall climbing robot scheme is introduced in detail.Aiming at the shortcomings of existing traditional machine vision ship waterline detection,this paper proposes a combination of segmentation and edge detection to remove the interference of the full load line according to the problems in practical application.The edge chain fusion algorithm Msedge effectively removes other interference lines.To make waterline detection easier.Traditional machine vision algorithms are artificial design features,and a single algorithm cannot handle all complex and varied environments accurately.This paper firstly proposes a deep learning algorithm to solve the problem of ship waterline detection.Based on the advanced real-time deep learning target detection network,a semantic segmentation network structure called Fine-Mobilenet is proposed,which makes the detection accuracy and robustness of the system both have great improvement.According to the characteristics of the problem to be solved,this paper analyzes and contrasts different dataset labeling schemes,and designs a set of efficient image annotation tools.In this experiment,a total of 30 videos of different vessels in different environments were collected in Qinhuangdao Port.Each video was marked with 200 to 500 images according to the situation.A total of 10,000 training sets were marked,and the final test accuracy reached 98%.Through the field test in Qinhuangdao Port,the rationality of the system and the accuracy of the test results were verified.