Image Segmentation For SAR Oil Slicks Based On Visual Attention Mechanism And Lightweight FCN

Image segmentation for SAR oil slicks can provide critical data for loss assessment and subsequent claims.In recent years,many studies have begun to use convolutional neural network for image segmentation of SAR images,and achieved good results.However,there are still some problems when these technologies are applied to practical problems.On the one hand,the memory requirement of convolutional neural network will increase with the increase of the size of the input,and the size of SAR images itself are generally very large.In practical applications,the cost of related hardware is obviously an important factor to be considered;If we use manual clipping as preprocessing,it will lead to the decline of the ease of use of the whole system.On the other hand,due to the limited data sources of oil spill accidents,the data set involved in this thesis has been expanded by using a variety of means,which leads to the image of the data set is clearer than the original image.This will lead to a decline of the accuracy of the deep learning model when processing the original images.To solve the above problems,the work of this thesis is as follows:Firstly,in order to realize the end-to-end process of the segmentation of SAR images in original size,this thesis designs an auxiliary algorithm for end-to-end segmentation of SAR images of oil slicks called visual attention algorithm,which is combined whit the related concepts of visual attention mechanism and its application cases in the field of image processing.This algorithm can assist the deep learning model to process the input image of any size,and effectively limit the use of memory.Making it possible to process the super-large SAR images of oil slicks with cheap equipment.Then,aiming at the problem of precision degradation caused by actual blurred image.In the further experiments,we find that the decline of accuracy is larger in the model with more parameters and lower in the model with less parameters,while the model with too few parameters and too simple structure still performs poorly in accuracy.Based on the framework of SegNet used by previous work,a new lightweight FCN model is designed by reducing parameters and adjusting the structure.The design concept of this model is to reduce the use of parameters as much as possible while maintaining the integrity of the overall structure,so as to reduce the decline in accuracy when dealing with blurred images.Finally,the visual attention algorithm and lightweight FCN model designed in this thesis are experimented and analyzed.Compared with other models,our model achieves a mIoU of 94.5% and 89.44% in processing normal images and blurred images;The model size and the amount of floating-point computation are also at a low level.In the overall test,we split an input image of 6281 × 3920 in about 2.65 seconds using only 3807 MB of memory.