Oil Spill Segmentation In Inland Rivers Based On ICNet

With the continuous development of production technology and industrial level,the social demand for various energy sources is increasing,uch as:oil resources.The transportation of oil resources mainly depends on waterways,Whether it is via offshore oil or inland water transportation(inland rivers),traffic accidents or other accidents inevitable occur,leading to oil spills and water pollution.Although,with the development of science and technology,the inland river transportation system is becoming more and more perfect,but it is followed by problems such as busier transportation routes and frequent traffic accidents resulting in oil spills and other serious pollution to water bodies.At present,the domestic and foreign oil spill segmentation technology has been relatively mature,about the division of oil spill area in the river is still slightly insufficient.In addition,the inland river environment is more complex than the sea surface environment,and there are more interference factors that make inland river oil spills more challenging.This paper takes the segmentation of 'inland river oil spill as the research theme and the specific work are summarized as follows:Firstly,due to the poor quality of inland river oil spill images caused by illumination factors,which cannot train the best segmentation model and affects the segmentation accuracy,an image restoration algorithm is proposed to improve the image quality.The reflection on the water surface is easy to be caused by the strong light.The image information of the light spot area caused by the reflection cannot be distinguished as water or oil.The characteristics of the oil spill area will not be obvious and the segmentation is difficult.In order to solve the lighting problem and improve the image quality,this paper proposes a preprocessing method of ACE image enhancement and FMM image restoration.A super-pixel block is added to the preprocessing method.The method of using pixel block as unit to repair the image instead of pixel point as unit in the original algorithm can accelerate the speed of image preprocessing.The ACE algorithm can effectively enhance the color characteristics of the image highlighting the oil spill,and the FMM algorithm can repair the information loss caused by the bright spots reflected by the light.Adding superpixels can speed up the image preprocessing.Secondly,the traditional segmentation method is too complexalg to achieve real-time segmentation.This paper proposes an ICNet network with Xception as the backbone to segment inland oil spills in order to solve the speed problem.The traditional segmentation method is too complex to achieve real-time segmentation.This paper proposes an ICNet network with Xception as the backbone to segment inland oil spills in order to solve the speed problem.Using pre-processed images as ICNet training data can effectively improve the accuracy of segmentation.Xception is used as the backbone network,deep separable convolution instead of ordinary convolution can reduce the amount of calculation,increase the speed of model calculation,and achieve real-time segmentation.Finally,the segmentation model is trained based on thousands of oil spill images.This method is compared with the classic segmentation algorithm(FCN,SegNet)and the lightweight network(UN et).The experimental results show that the comprehensive performance of the proposed method is better than the other three models.