Development Of Port Oil Spill Identification And Location System Based On Deep Learning

In recent years,the continuous development of marine transportation,oil development and other industries has brought new vitality to economic growth,but also promoted the frequent occurrence of marine oil spills.As a trading center,the frequency of oil spill accidents is not optimistic.Because of the unpredictability of the oil spill accident,and the oil diffusion is extremely rapid,which will cause great harm.It is an important problem to quickly identify the oil spill and extract the important positioning information,which can make the follow-up oil containment work more accurate and efficient,and prevent the secondary spread of oil spill,which is a major problem to be solved.Based on the oil spill recognition model of YOLOv3(You Only Look Once:version 3)and the edge detection method of contour extraction function,this paper mainly completes the following parts:(1)Several currently widely used recognition algorithms based on Convolutional Neural Network(CNN)are analyzed.Based on the characteristics of oil spill,YOLOv3 model is selected as the basic framework for oil spill recognition.An oil spill recognition data set based on UAV images and videos is constructed.(2)Based on the YOLOv3 framework,an oil spill recognition model is built.The default backbone network Darknet-53 of the YOLOv3 model is replaced by the MobileNetv1 to solve the network redundancy problem when it’s used to recognize a single and large area target,and the recognition speed is improved.Multiple prediction frame sizes are generated by K-means clustering algorithm,and the optimal number of prediction frames is determined by analyzing the influence of the number of sizes on the prediction accuracy.The idea of GIOU is combined with the Loss function to solve the problem of limited conversion between IOU and Loss function.The Non-maximum suppression(NMS)algorithm is replaced by the Soft-NMS algorithm to improve the shortcomings of the NMS algorithm in the process of anchor filtering,and the recognition accuracy is improved.(3)Grayscale,mean filtering,binarization and other image preprocessing areperformed on the oil spill image.Canny edge detection algorithm and OpenCV contour extraction function are used to accomplish the edge detection on the image previously preprocessed.Base on the results of the two methods,the contour extraction function is selected as the edge detection method,and the location of the oil spill edge is generated.The distances between the reference point and all edge points are calculated to select the farthest distance of the oil spill edge and the coordinates of the corresponding point.(4)The GUI interface of oil spill identification system is developed,and the test experiment is designed to verify the running effect of the program.