| Water gauge weighing is a common method of ship deadweight statistics.This method of weighing is not only quick and convenient,but also scientific and accurate.It is widely used in the shipping industry both at home and abroad.At present,the main method of measuring ship water gauge readings is manual visual inspection,however,manual visual inspection has problems such as excessive error and difficult to record and archive.In recent years,with the rapid advancement of computer computing power,there have been domestic researches on the recognition of ship water gauge readings in the field of digital image processing.However,most of the recognition methods need a harsh environment,not only need to collect high-quality image data,but also the recognition process is overly dependent on manual adjustment of algorithm parameters.Therefore,these recognition methods are not suitable for complex environments in natural scenes.In response to the above problems,this paper researched and designed a water gauge reading recognition system based on deep learning and machine vision.The system performs a large number of recognition experiments on the video collected on site to ensure that the final water gauge reading has a certain degree of accuracy and scientificity.This article mainly studies the following aspects:(1)The overall scheme of the system in this article is divided into a hardware part and a software part.The hardware part is used for image acquisition and software support.It mainly includes an image resource acquisition unit,a handheld terminal console and a notebook computer.Each module communicates through LAN.The image resource collection unit has two methods: for the situation where the water gauge mark on the back shore cannot be directly observed,it is collected by a drone;for the situation where it is difficult to collect the water gauge mark on the side of the shore,I designed a handheld image acquisition device.The handheld image acquisition device can adjust the shooting angle of the shooting equipment at multiple angles between the narrow shore roads.The handheld terminal console controls the shooting of the shooting equipment.After the captured video is confirmed by the staff,the video is sent to the laptop via LAN.(2)Research on waterline detection algorithmAiming at the inaccurate waterline detection caused by uneven image illumination,water surface highlights and water trace interference,this paper proposes an improved UNet network.In order to obtain more feature information and improve the accuracy of the network,the Efficient Net-B0 network obtained by optimizing the depth,width and resolution of the Mnas Net network is used as the coding part of the improved UNet network to perform feature extraction on the image.At the same time,in order to reduce the information loss of the network in the process of convolution and pooling,this paper is based on the feature fusion method of the UNet network,and adds the fusion of deep features and shallow features to the decoding part of the improved UNet network.Finally,the paper analyzes the waterline detection capability of the improved UNet network through a large number of experiments,and the overall detection success rate of the improved UNet network waterline is 93.8%.(3)Water gauge reading recognitionPerforming MSER extraction on grayscale images will cause the water gauge mark to be missed due to the low brightness of the image.To solve this problem,this article adds the MSER extraction in the L channel of the CIELab color space.The water gauge reading recognition process is as follows: First,according to the contact between the extracted water gauge mark candidate area and the waterline,the water gauge mark candidate area is divided into the candidate connected area at the non-waterline and the candidate connected area at the waterline;secondly,Filter the BLOB feature of the water gauge mark and the classification and recognition of the convolutional neural network for the candidate area at the non-waterline;then,calculate the actual reading of the single-frame water gauge image according to the geometric characteristics of the connected area of the water gauge mark at the waterline;finally,Find the average of the two values based on the maximum and minimum readings of the sequence image recognition,that is,the final reading of the video.This paper analyzes the accuracy of software recognition through a large number of video tests.The experimental results show that the algorithm has an accuracy of 96.2% for videos with good quality,and it is also more accurate for videos with poor quality due to collection reasons.The recognition accuracy rate of videos with poor shooting quality can reach 82.6%.(4)Water gauge recognition visualization systemIn order to facilitate the operation of the staff,this article uses the PyQt5 framework to build the visual interface on the PyCharm development platform.The system interface has three modes,namely real-time input mode,video input mode and accuracy rate statistics mode.Before the recognition process,the staff can crop the video on the system.After the recognition is over,the interface will display the 6 recognized image readings and save the recognition data at the same time. |
