DSF-Net: A Dual Feature Shuffle Guided Multi-Field Fusion Network for SAR Small Ship Target Detection

Abstract

SAR images play a crucial role in ship detection across diverse scenarios due to their all-day, all-weather characteristics. However, detecting SAR ship targets poses inherent challenges due to their small sizes, complex backgrounds, and dense ship scenes. Consequently, instances of missed detection and false detection are common issues. To address these challenges, we propose the DSF-Net, a novel framework specifically designed to enhance small SAR ship detection performance. Within this framework, we introduce the Pixel-wise Shuffle Attention module (PWSA) as a pivotal step to strengthen the feature extraction capability. To enhance long-range dependencies and facilitate information communication between channels, we propose a Non-Local Shuffle Attention (NLSA) module. Moreover, NLSA ensures the stability of the feature transfer structure and effectively addresses the issue of missed detection for small-sized targets. Secondly, we introduce a novel Triple Receptive Field-Spatial Pyramid Pooling (TRF-SPP) module designed to mitigate the issue of false detection in complex scenes stemming from inadequate contextual information. Lastly, we propose the R-tradeoff loss to augment the detection capability for small targets, expedite training convergence, and fortify resistance against false detection. Quantitative validation and qualitative visualization experiments are conducted to substantiate the proposed assumption of structural stability and evaluate the effectiveness of the proposed modules. On the LS-SSDDv1.0 dataset, the mAP50−95 demonstrates a remarkable improvement of 8.5% compared to the baseline model. The F1 score exhibits a notable enhancement of 6.9%, surpassing the performance of advanced target detection methods such as YOLO V8.

A Complex Background SAR Ship Target Detection Method Based on Fusion Tensor and Cross-Domain Adversarial Learning

“…Combining images of different modalities and participating in training is known as multimodal fusion of images [30][31][32].…”

“…This approach enabled deep fusion between different blocks. Xu et al [31], proposed the GWFEF-Net, a group feature enhancement and fusion network with bipolar feature enrichment, for improved bipolar SAR ship detection. Zhang et al [32], introduced SuperYOLO, an accurate and fast RSI target detection method that fuses multimodal data and employs assisted super-resolution (SR) learning for high-resolution (HR) target detection of multi-scale targets.…”

The Analysis of Intelligent Functions Required for Inland Ships

“…Thompson et al [13] proposed an efficient LiDAR-based target segmentation method for the marine environment that utilizes 3D occupancy grid segmentation to effectively map large areas. Xu et al [14] proposed a novel network architecture for small SAR ship target feature extraction and multi-field feature fusion combined with dual-feature mobile processing based on bridge node and feature assumptions, which solved the problem of misdetections and false detections in the detection of small SAR ship targets. Ye et al [15] proposed an EA-YOLOv4 algorithm with an augmented attention mechanism, which utilizes a convolutional block attention module (CBAM) to search for features in the channel dimension and spatial dimension, respectively, to improve the feature perception ability of the model for ship targets.…”

Swin-PAFF: A SAR Ship Detection Network with Contextual Cross-Information Fusion

“…Although Swin-PAFF combines Transformer and CNN detection methods, we intend to explore ways to reduce the computational complexity associated with the Transformer model by integrating it with blockchain in future research [43,44]. In addition, we plan to explore ways to make the Transformer model lighter.…”