BBox-Free SAR Ship Instance Segmentation Method Based on Gaussian Heatmap

Recently,  deep  learning  methods  have  been  widelyadopted  for  ship  detection  in  synthetic  aperture  radar  (SAR)images.  However,  many  of  the  existing  methods  miss  adjacentship  instances  when  detecting  densely  arranged  ship  targets  ininshore scenes. Besides, they suffer from the lack of precision inthe instance indication information and the confusion of multipleinstances  by  a  single  mask  head.In  this  paper,  we  propose  anovel center point prediction algorithm, which detects the centerpoints  by  finding  a  long  distance  variation  relationship  betweentwo points. The whole prediction process is anchor-free and doesnot require additional bounding box (BBox) predictions for non-maximum suppression (NMS). Therefore, our algorithm is BBox-free and NMS-free, solving the problem oflow recall rateswhenconducting  NMS  for  densely  arranged  targets.  Furthermore,to  tackle  the  deficiency  of  position  indication  information  inlocalization  tasks,  we  introduce  a  feature  fusion  module  withfeature  decoupling  (FD).  This  module  uses  classification  branchto  provide  guidance  information  for  localization  branch,  whilesuppressing the influence of the gradient flow mixing, effectivelyimproving  the  algorithm’s  segmentation  performance  of  shipcontours.  Finally,  through  principal  component  analysis  (PCA)of   the   Gaussian   distribution   covariance   matrix,   we   proposea  loss  function  based  on  the  distance  between  centroids  andthe  difference  of  angle,  called  centroid  and  angle  constraint(CAC).  CAC  guides  the  network  in  learning  the  criterion  thata  single  dynamic  mask  head  is  only  valid  for  a  single  instance.Experiments conducted on PSeg-SSDD and HRSID demonstratethe effectiveness and robustness of our method.