基于改进UNet网络的船舶水尺读数识别方法研究

Research on recognition method of ship water gauge reading basedon improved UNet network

针对船舶水尺识别过程中出现的图像亮度不均、高光、水尺标志倾斜和轻微模糊粘连以及船体表面划痕和水尺标志损坏等,本文提出一种基于改进UNet网络的船舶水尺识别方法。首先,利用改进的UNet网络对图像进行吃水线检测﹐从图像中分离船舶区域和水面区域;其次,在分离后的船舶区域图像内,通过在多通道色彩空间图像上提取各图像的最大稳定极值区域MS-ER,将各图像最大稳定极值区域并集操作的结果作为水尺标志候选区域,再把水尺标志候选区域分为非吃水线处候选连通区域和吃水线处候选连通区域;然后,基于水尺标志连通区域的blob特征对非吃水线处水尺标志连通区域初次筛选,再利用卷积神经网络对筛选后的连通区域所在的ROI图像进行分类识别;最后﹐根据吃水线处水尺标志连通区域的几何特征计算水尺实际读数。实验结果表明,该算法对质量较好的图像有较高的识别准确率,准确率可以达到96.8%,对于采集﹑船体等原因导致质量较差的图像也能有80.7%的准确率。

We have proposed a method for identifying water gauges on ships,aiming at the uneven brightness of images,highlights,tilting,blurring and sticking of water gauge marks,as well as the scratches onthe hull surface and damage to water gauge marks during the recognition process.Firstly,we have usedthe improved UNet network to detect the waterline of the image and separate the part of ship and thepart of water from the image.Secondly,we have extracted the candidate connected regions of watergauge mark through MSER in the multi-channel color space of the separated ship part image,and dividedthe candidate regions of water gauge mark into non-waterline candidate regions and waterline candidateregions.Thirdly,we have filtered the non-waterline candidate regions through their BLOB and classifiedthem with CNN.Finally,Finally,combined with the water gauge mark at the non-waterline that has beenidentified,the actual reading of the water gauge mark is calculated through the filtering and geometriccharacteristics of the connected area at the waterline,Experimental results show that the algorithm has ahigher recognition accuracy rate for images with better quality,and the accuracy rate can reach 96.8%,and it can also have an accuracy rate of 80.7%for images with poor quality due to collection,ship hulland other reasons.