Anchor-Free SAR Ship Instance Segmentation With Centroid-Distance Based Loss

Abstract

Instance segmentation methods for synthetic aperture radar (SAR) ship imaging have certain unsolved problems:(1) Most of the anchor-based detection algorithms encounter difficulties in tuning the anchor-related parameters and high computational costs. (2) Different tasks share the same features without considering the differences between tasks, leading to mismatching of the shared features and inconsistent training targets. (3) Common loss functions for instance segmentation cannot effectively distinguish the positional relationships between ships with the same degree of overlap. In order to alleviate these problems, we first adopt a lightweight feature extractor and an anchor-free convolutional network, which effectively help to reduce computational consumption and model complexity. Secondly, to fully disseminate feature information, a dynamic encoder-decoder is proposed to dynamically transform the shared features to task-specific features in channel and spatial dimensions. Thirdly, a novel loss function based on centroid distance is designed to make full use of the geometrical shape and positional relationship between SAR ship targets. In order to better extract features from SAR images in complex scenes, we further propose the dilated convolution enhancement module (DEM), which utilizes multiple receptive fields to take full advantage of the shallow feature information. Experiments conducted on the SAR Ship Detection Dataset (SSDD) prove that the method proposed in this paper is superior to the other state-of-the-art algorithms in terms of instance segmentation accuracy and model complexity.

DCTC: Fast and Accurate Contour-Based Instance Segmentation With DCT Encoding for High-Resolution Remote Sensing Images

“…Comparison on the SSDD dataset: For SSDD dataset, models for SAR image instance segmentation specifically are chosen, such as C-SE Mask R-CNN [50], EMIN [51], FL-CSE-ROIE [52], MAI-SE-Net [53], SA R-CNN [54] and LFG-Net [6], as well as other classic methods. Table II exhibits the comparison results.…”

A Benchmark Sentinel-1 SAR Dataset for Airport Detection

“…It consists of YOLO v5s, YOLO v5m, YOLO v5l, and YOLO v5x network structures of different depths and widths. In the field of remote sensing, YOLO v5 is applied to crop circle detection [54] and vessel detection [55].…”

“…More regions of line Low values of D lead to inconsistent detections, while too large values of D will have the effect of overcutting the line segments into small. A detailed analysis of the four parameter settings for the line segment detection can be found in [55].…”

A Mask Attention Interaction and Scale Enhancement Network for SAR Ship Instance Segmentation

“…Zhao et al [16] designed a synergistic attention for better accuracy, but there were still many missed detections in complex scenes. Gao et al [17] proposed an anchor-free model with a centroid distance loss to enhance performance, but their models still lacks the capacity to deal with complex scenes and cases. Zhang et al [18] used the hybrid task cascade (HTC) [19] for SAR ship instance segmentation, yielding better performance than Mask R-CNN [20].…”