Research On Underwater Image Processing And Object Detection Algorithm Based On Nuclear Power Plant Cold Source Monitoring And Warning

In recent years,blockage of the water intake happened frequently in nuclear power plants.It caused loss of cold source of the nuclear reactor system,which directly affected the operation of the generating units and the safety of the reactor system.One of the most fundamental reasons for the cold source loss is the lack of intelligent realtime monitoring technology which could provide early warning of such blockage in the water intake of nuclear power plants.Therefore,the invasion of different types of marine organisms cannot be handled in time.Thus,based on the underwater image of the water intake of a nuclear power plant,this thesis firstly aims to analyze the results of applying different deep learning algorithms in underwater image processing and target classification and recognition;and then explores the feasibility of applying artificial intelligence,featured with real-time monitoring and warning technology,in the nuclear power plants.To fulfill the research goal,this thesis adopted artificial intelligence methods to realize real-time classification of marine organisms.The main contents of this thesis are as follows:(1)To obtain the underwater image of the water intake of the nuclear power plant,a new video collecting system has been investigated.Firstly,an underwater video collection system was constructed by using an improved waterproof camera;secondly,to filter and intercept images of sea creatures from a large amount of underwater videos,two inter-frame differential algorithms and image binarization algorithms were applied to the development of a program aimed to intercept images automatically by distinguishing the characteristics of sea creatures moving with the seawater in the monitored area.(2)Different image-enhancing algorithms has been studied and special datasets has been produced consequently using the images pre-processed,respectively.Image quality enhancing experiments were conducted,including traditional image enhancing algorithms such as image denoising,digital image enhancement,and the deep learningbased SRGAN(Super-Resolution Generative Adversarial Network)algorithm.To prepare for the subsequent training and validation of the model,three control datasets-FISH-2000(original images),DATASET-01(images processed by traditional image enhancing algorithm)and DATASET-02(images processed by SRGAN)have been constructed.(3)Different target detection algorithms were investigated,image recognition models were trained and analyzed consequently.Specifically,the work adjusted the input datasets,controlled the training parameters,and tested five target detection algorithms(SSD(Single Shot Muolti Box Detector),YOLO(You Only Look Once)v3,YOLO v4,YOLO v5,and YOLO X)which are capable of both detecting and classifying tasks.In doing so,the m AP(mean Average Precision),AP(Average Precision)and FPS(Flames Per Second)of the model obtained from the experiment have been calculated and analyzed.It was found that the SRGAN-YOLO X-640 model trained by using the YOLO X algorithm input dataset DATASET-02 gained the best recognition results with the AP(98.32%)and the FPS(27.76).(4)Finally,the best performing model has been improved and the feasibility of deploying the model for embedded platforms in practical use has been investigated.A different version of the lightweight neural network Mobile Net was introduced to replace the YOLO X algorithm’s backbone feature extraction network,with the purpose of reducing the number of parameters and the computational complexity in the network.After the training of the improved network,the computational speed of the new model was greatly improved at the expense of the decrease of recognition accuracy.The best recognition model(SRG-Mobile V3YOLO-640)had an FPS of 43.17,and an AP of only79.95%.