Research On Visual Detection Method Of Underwater Cables Based On Deep Learning

Underwater cables play an important role in the development of marine resources,the economic development of islands and the security of the national waterfront.Cable damage often bring inconvenience and loss to the marine industry and the lives of islanders.Therefore,real-time detection of underwater cables has become particularly important.Studies have shown that underwater cable detection usually uses the traditional visual detection method of underwater robots with cameras,but this method was costly which the detection accuracy and accuracy rate was low.Currently,deep learning-based visual detection methods have become the mainstream of target detection,and based on their advantages of high recognition accuracy,fast detection speed and high accuracy,a research on the visual detection method of underwater cables based on deep learning is carried out in this paper.The specific research contents are shown as following:Firstly,an in-depth analysis of the current situation of underwater image enhancement and target detection development at home and abroad was carried out.According to the current research status and problems of target detection technology,the target detection method was used deep learning is determined in this paper,and the related theory of deep learning was introduced to lay the foundation for the following technical research.Secondly,the initial dataset of underwater cables was established.Aiming at the problem that the number of underwater cable images were small and difficult to be obtained,based on the analysis of traditional data expansion,a progressive generation countermeasure network was introduced and the network model was improved to generate high-quality cable images,and the expansion of initial data was completed in order to reduce the impact of image generation by traditional data expansion on deep learning training.For the image quality problems such as image color bias and low contrast brought by underwater environment,the underwater imaging principle was analyzed,and a multi-method fusion underwater image enhancement algorithm was proposed for image quality improvement to provide high quality data images for target detection.Thirdly,the structure,advantages and disadvantages of feature extraction network and mainstream deep learning algorithm were introduced.According to the requirements of accuracy,running time and portability of underwater cable target detection algorithm,YOLO v4 was selected as the basic target detection network which will be improved in feature extraction network and loss function to realize the two-dimensional detection of underwater cable.Through comparative experiments,the advantages of the improved network model in detection speed and accuracy were verified.At the same time,it was shown that underwater image quality enhancement can be effectively improved.Finally,the binocular camera model and imaging principle were briefly introduced,the internal and external camera parameters were solved by binocular camera calibration,and the aberrations were eliminated by polar line correction.Based on the commonly used feature extraction algorithm and feature matching algorithm,the search for feature matching was reduced by identifying regions of interest for deep learning underwater cable detection.Using the ORB feature point extraction algorithm and Flann Based Mather feature matching algorithm suitable for underwater cable image,the three-dimensional positioning of underwater cable was realized,and the speed and accuracy of target positioning were improved.