Research And Application Of Underwater Character Recognition Algorithm Based On Deep Learning

In recent years,with the development of marine resources,underwater visual technology has continuously improved.At the same time,underwater character recognition has also attracted scholars’ attention with the increasing number of underwater robot operations.The rapid development of visual technologies such as deep learning and image processing has led to the research and application of Optical Character Recognition(OCR)technology in various scenarios by many scholars.Firstly,traditional character recognition algorithms are only suitable for easily recognizable scenarios,such as concise and clear fonts,high contrast between characters and background,and relatively simple backgrounds.Faced with the research on character information recognition methods in complex natural backgrounds such as underwater character recognition,the preprocessing algorithm of underwater original images has also become a research hotspot.Secondly,difficulties such as blurry contrast,distorted colors,and complex backgrounds in underwater videos or images make character recognition inefficient,posing challenges to character recognition technology.This thesis focuses on two main directions: image enhancement and character detection and recognition in underwater scenes,in order to achieve an accurate and efficient algorithm application for underwater character recognition and assist in the inspection of underwater mechanical part numbers.The thesis work of this thesis mainly includes the following three aspects:Image enhancement is the first step in character recognition in complex underwater environments.To solve the problem of image degradation caused by underwater imaging,three image enhancement algorithms are compared: histogram equalization algorithm,SSR algorithm based on Retinex model and dark channel prior defogging algorithm.Finally,the dark channel prior defogging algorithm is selected as the underwater image pre-processing algorithm in this paper.By compensating for the red channel and white balance,color correction is performed and a dark channel image is obtained again.Then,the background light value and transmittance are estimated to obtain the algorithm enhanced image in this paper.This thesis improves the algorithm for image enhancement on the UIEBD dataset and achieves better results in PSNR,SSIM,and UCIQE image enhancement evaluation standards.At the same time,the output image has natural colors and increased contrast,which can meet the needs of underwater character recognition.Character recognition is the core step of underwater character recognition.Firstly,by comparing the existing two-stage and end-to-end classic character recognition algorithms,the more efficient end-to-end FOTS algorithm was selected for character recognition.Then,the FOTS algorithm is improved to solve the two problems of insufficient Receptive field and unbalanced sample training.First,the ASPP structure is added to the FOTS algorithm feature extraction network to enhance the algorithm Receptive field;Secondly,the standard cross entropy loss function of the algorithm is replaced and improved to Focal loss function to increase the training weight of difficult positive samples.This thesis improves the algorithm and improves the end-to-end character recognition accuracy from 83.55% to 85.80% in the ICDAR 2015 public dataset.Finally,an accurate and efficient underwater character recognition algorithm was implemented to assist in the inspection of underwater mechanical parts.Firstly,a dataset containing mechanical part numbers with various underwater degradation features was created.Afterwards,an underwater character recognition algorithm was designed that includes underwater image enhancement and character recognition,and algorithm optimization was carried out for images in the same underwater environment.Finally,visualize the identification results of the numbers through software display.