卢鹏涛,蒋雯,黄菊,孙水发,汪方毅.基于文本图像修正的两阶段船名识别框架[J].电子测量与仪器学报,2024,38(2):30-39 |
基于文本图像修正的两阶段船名识别框架 |
Two-stage vessel name recognition framework based on text image correction |
|
DOI: |
中文关键词: 船舶名称识别 文本检测 透视变换 超分辨率 |
英文关键词:vessel name recognition text detection perspective transform super-resolution |
基金项目:国家自然科学基金(61871258)项目资助 |
|
Author | Institution |
Lu Pengtao | 1.Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering, China Three Gorges University,Yichang 443002,China; 2.College of Computer and Information Technology, China Three Gorges University, Yichang 443002,China |
Jiang Wen | Yangtze Yichang Communications Authority, Yichang 443001,China |
Huang Ju | 1.Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering, China Three Gorges University,Yichang 443002,China; 2.College of Computer and Information Technology, China Three Gorges University, Yichang 443002,China |
Sun Shuifa | Hangzhou Normal University, Hangzhou 311121,China |
Wang Fangyi | 1.Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering, China Three Gorges University,Yichang 443002,China; 2.College of Computer and Information Technology, China Three Gorges University, Yichang 443002,China |
|
摘要点击次数: 375 |
全文下载次数: 555 |
中文摘要: |
船舶名称(牌照)识别在水路运输系统中发挥着重要的作用。针对船舶名称在内河航道中目标较小且航道两岸观测船舶存在较大倾斜角度导致难以识别的问题,提出一个以自然场景文本检测算法(differentiable binarization, DB)和文本识别算法(convolutional recurrent neural network, CRNN)为基础的船舶名称自动识别框架(automatic ship name identification,ASNI),ASNI包括以下3个部分:船名检测、文本图像修正和识别,其中,船名文本图像修正由船名矫正模块和超分辨率重建模块构成。首先,该框架利用DB算法对图像船名候补区域特征进行自适应尺度融合处理获取特征图,通过特征映射预测生成的二值图像寻找连接区域,以此获得船名感兴趣区域(ROI)。其次,在船名检测之后引入船名矫正模块,基于透视变换对ROI中船名不规则文本进行矫正。此外,设计超分辨率重建模块,对矫正后的船名图像进行超分辨率重建处理,以提高船名图像的分辨率。最后,利用CRNN算法对文本图像修正后的ROI中船名进行识别得到最终结果。通过在内河航道船舶数据集(ship license plate,SLP)上进行训练和测试,最终实验结果显示,ASNI框架对船舶识别的平均准确率为87.50%,相比于基础框架提升了3.12%。本文设计的框架有效解决了因分辨率不足和倾斜导致船舶识别不准确的问题,相比基础框架,ASNI有更好的识别效果。 |
英文摘要: |
The recognition of vessel names (license plates) plays a crucial role in waterway transportation systems. Addressing the challenge of identifying vessel names in inland waterways, where targets are relatively small and vessels are observed at significant angular inclinations on both sides of the waterway, we propose an automatic ship name identification (ASNI) framework based on the differentiable binarization (DB) natural scene text detection algorithm and the convolutional recurrent neural network (CRNN) text recognition algorithm. ASNI comprises three main components: ship name detection, text image correction, and recognition. The text image correction component consists of a ship name correction module and a super-resolution reconstruction module. Firstly, the framework utilizes the DB algorithm to perform adaptive scale fusion processing on the candidate regions of vessel names in images, generating feature maps. Feature mapping is used to predict and generate binary images to identify connected regions, thereby obtaining regions of interest (ROI) containing vessel names. Subsequently, after ship name detection, a ship name correction module is introduced to rectify irregular text within the ROI using perspective transformation. Furthermore, a super-resolution reconstruction module is designed to enhance the resolution of the corrected vessel name images. Finally, the CRNN algorithm is employed to recognize vessel names within the corrected text images in the ROI, yielding the ultimate results. Through training and testing on the ship license plate (SLP) dataset specific to inland waterways, experimental results demonstrate that the ASNI framework achieves an average accuracy of 87.50% in vessel recognition, representing a 3.12% improvement over the baseline framework. The framework presented in this paper effectively addresses issues related to low resolution and angular inclinations leading to inaccurate vessel recognition. Compared to the baseline framework, ASNI exhibits superior recognition performance. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|