Research On Ship Number Text Recognition Based On CRNN Model Based On Joint Image Restoratio

With the rapid development of the economy,water transportation is becoming increasingly busy.Therefore,it is particularly important to do a good job in ship detection and identification,and achieve comprehensive management of water transportation logistics.At present,ship number detection and recognition technology has also been widely applied by regulatory authorities in various river checkpoints.However,due to the problems of low recognition accuracy and large recognition delay in traditional ship recognition technology,there are certain limitations in practical applications.With the development of deep learning theory,character recognition based on deep features has become a focus of current research.However,due to text occlusion and text blurring,it is difficult to obtain ideal recognition results based on deep learning models.Although the recognition rate can be improved by increasing the training set,it is difficult to obtain samples suitable for ship side text and the calibration cost is high.In response to these issues,this article proposes a joint image restoration technology to improve text recognition rate,thereby providing effective quantitative support for regulatory authorities.The specific work is as follows:This proposes a joint image restoration and recognition network to address the issues of text blur and occlusion.Laplace operator is used to determine whether the current text is ambiguous,and CycleGAN is used to restore it,so as to improve the robustness of the model to text ambiguity.Use GAN discriminator loss to determine whether there is text occlusion in the current text,and use Pix2pix model to restore the occlusion area,thereby reducing the impact of text occlusion.The dual branch network proposed in this article for joint image restoration can achieve ideal analysis results in situations of strong text blur or occlusion.The experimental results show that the method proposed in this thesis can meet the real-time supervision requirements of navigation channels.For the small sample problem,this thesis enhances the data set,including translation,rotation,scale scaling,etc.,to simulate the ship status in different scenes,and adjusts the brightness to meet the ship status under different conditions,thus reducing the risk of model overfitting.The experimental results on the CTW dataset and the real ship chart dataset show that the model proposed in this thesis has significantly improved the accuracy of fuzzy text recognition and occluded text recognition,providing strong data support for regulatory authorities.