Research On Image Generation Of Surface Oil Spill And Its Detection Methods Based Deep Learning

With the continuous exploitation of land resources by various countries in recent years,the natural resources on land have been plundered away.Many countries began to exploit the vast ocean resources,although frequent exploitation has brought more resources,but the impact on the oceans.Oil spill pollution is one of the most serious of marine pollution,oil spill pollution from oil spill accidents has brought terrible influence on marine ecology.Therefore,it is very necessary to carry out timely and rapid inspection in the event of an oil spill accident,and the research of marine oil spill inspection has an important role to play.This topic focuses on the task of oil spill detection under visible light cameras,which are fast and convenient to deploy,but under visible light cameras,the amount of available oil spill data is insufficient,and the color characteristics of oil spill areas under visible light are different,which is difficult to detect.To solve the problem of small amount of oil spill data,this paper uses image structure network and image texture generation network to generate new oil spill data for the first problem.In second problem that oil spill area is difficult to detect in the visible light,this paper proposes a new oil spill area segmentation algorithm based on deep learning.The specific contents of this paper are as follows.For the task of oil spill image structure generation,the image generation is mainly performed by generating adversarial network.Firstly,a controlled Gaussian fuzzy module is added in front of the network input to expand the data set of the original network.Secondly,the two-way ResBlock structure is used to solve the problem that the network training is prone to crash.Finally,the data with similar morphology as the oil spill region is fed into the trained network to obtain new data with the edge structure of the oil spill region.For the oil spill image texture generation task,the generative adversarial network is also utilized for texture generation.In this task,the output of the structure generation task is used as the input of the texture generation task,and the oil spill is performed using the idea of the map transformation task texture bonding of regions.The new normalization layer is designed to join the original generator network structure to produce a sharper oil spill image,which results in the new reliable oil spill images to augment the original oil spill dataset.For the task of oil spill image segmentation,this paper proposes a dual segmentation network to solve the problem of difficult detection due to different color characteristics in visible light.The dual segmentation network is divided into a pixelwise segmentation network and a patch-level segmentation network,and the pixelwise segmentation network generate pixelwise segmentation result of grayscale oil images.Patch-level segmentation task divides the original RGB oil spill image into several pixel patches and classifies these blocks.Finally,the segmentation results of the two networks are fused to obtain the final oil spill region.