Semantic Segmentation Of Ship Image Based On Visible Light Remote Sensing

With the rapid increase of population and international competition,land resources are becoming more and more scarce,so that the ocean becomes an important way for coastal countries to promote their own economic development and relieve the space pressure nowadays.Ships are effective tool for the development and utilization of ocean resources,for which the detection and control of them are becoming more and more important.As a hot spot in the field of satellite remote sensing detection,the detection technology based on visible light remote sensing image has a good effect for ships on the ocean.Among them,the recognition technology based on convolutional neural network has better robustness and accuracy than the traditional one.Therefore,this paper studies the semantic segmentation of ship image based on convolution neural network.Firstly,the image data were collected.The original images of the data set were obtained from three different places,after marking the original images without label,removing damaged images,deleting images with low resolution,and balancing the positive and negative data sets,the images were finally enhanced to reduce the probability of over-fitting during the training process.Secondly,the coding layer network structure of semantic segmentation of visible light remote sensing ship image was designed,four main modules in which were: res_block,incp_block,shuffa_block,and shuffb_block.And then these four modules were integrated to introduce the network structure of the whole coding layer.The characteristics of the network structure were as follows:(1)in addition to using traditional convolution,the whole model also used a large number of depth separable convolutions,the parameters of which were smaller than that of traditional convolutions,so it had a faster calculation speed;(2)residual connection was used in the res_block,and the flow of gradient was enhanced through skip layer connection,which also prevented the network from degradation with the deepening of model training depth.(3)The convolution layers with different convolution kernels were used for parallel convolution in the incp_block,which was better for collecting the features of ships in the remote sensing images and making dense connection and fusion of different feature data;(4)the group convolution was used in the shuffa_block,which could speed up the operation.(5)SE attention module was used in the model to improve the weight of target features,and then improved the final accuracy of the model.The network structure of the coding layer was used to experiment on the data set processed before and compared with other network structures to ensure it had a good effect and efficiency for the classification of visible remote sensing satellite ship images.