| With the increasing demand for Marine resources,how to safely and effectively exploit and utilize Marine resources in the complex and changeable Marine environment is still a research focus of Marine exploration and utilization.As a big maritime country,China has a long coastline and vast sea area,rich in Marine resources,but also faces many Marine disputes.As one of the basic tasks of Marine exploration,the visual-based sea surface environment perception technology plays an important role in the development and utilization of Marine resources and sea navigation.In this paper,the sea surface environment is taken as the research background,and the image pretreatment technology is used to process the visible images of sea surface objects in the early stage.Based on the theoretical framework of deep learning,the semantic segmentation model of sea surface objects is designed by using void convolution and spatial pyramid pooling structure,and the model is optimized by expanding the network depth.The specific research contents are as follows:Firstly,the surface image preprocessing,complex surface water environment humidity,sea fog,light conditions,the formation of natural factors such as interference,make visible light sensor imaging blur,the edge of the area part of semantic information and do not distinguish clearly,at the same time in the area of the surface may also be because of water splashing up and the reflected wave,etc.,the existence of distractors,The continuity of the whole semantic area is destroyed,and the reflection of objects on the water surface causes the boundary between the water bank and the water surface is not obvious,which will cause the image blur and noise problems.Therefore,it is necessary to carry out filtering and enhancement algorithm research on sea surface visible image,aiming to recover the lost texture and edge change information in the image,and at the same time to preserve more information and not to produce loss in the subsequent processing.Secondly,based on the weak supervision and learning network to visible light image boundary exploration,due to the water quantity is less,the target data set training model did not have enough quantity,the model can’t from the training process to extract target features,in such a situation to fit the data,will lead to loss function while training performance is good,but the prediction errors is higher.In this paper,we demonstrate a simple and effective weakly supervised learning method based on the idea of explicitly exploring target boundaries from training images to maintain consistency of segmentation and boundaries.The obtained rough graph is used to synthesize boundary annotations and the annotations are used for training.The network further mines more object boundaries and provides constraints for segmentation.Hole in the end,based on the convolution of the surface of visible light image target semantic segmentation algorithm research,it has to do with the visible light image characteristics highly correlated,this paper designs a hole based on convolution surface targets of semantic network segmentation,by deepen the depth of the network,expand the empty space in the pyramid structure of pooling branch module,realize the semantic segmentation model of response to the surface of the scene,The training and prediction are carried out on the self-made sea surface visible image dataset,and a better segmentation effect is achieved. |
