Improved YOLOX’s Anchor-Free SAR Image Ship Target Detection

Abstract

For the characteristics of synthetic aperture radar (SAR) images, such as the dense arrangement of ship targets on shore, which are easily affected by land, the sparse distribution of small ships in the deep sea, which are easily missing detect, and also the existence of a lot of negative sample background areas. We propose a new method based on improved YOLOX as Anchor Free target detection method, which greatly improves the training efficiency compared with the preset anchor box. In view of the problem that densely packed ship targets are easy to miss, and the ship target in the deep sea is weak and the distribution is sparse. We propose improved corner efficient intersection over union (ICEIOU) to further comprehensively consider the regression parameters, and the loss function is optimized for the network training process. We use Adaptive-NMS to adaptively adjusts the non-maximum suppression (NMS) threshold value for the dense arrangement and sparse distribution of ships, and atrous convolution is used to improve the problem of some detailed information loss. And combined with coordinate attention mechanism, high-speed and high-precision ship target detection is realized. The experimental results show that compared with the original YOLOX method, the detection precision of the method in this paper on HRSID dataset is improved from the original 88.81% to 91.48%, and the mAP index is improved from the original 87.72% to 91.76%, which is obviously better than the comparison method.

CCDN-DETR: A Detection Transformer Based on Constrained Contrast Denoising for Multi-Class Synthetic Aperture Radar Object Detection

“…Additionally, integrating SAR-based prior knowledge [ 31 , 32 ] with deep learning helps to provide more stable detection results. Moreover, some specifically optimized loss functions [ 12 , 13 ] have also gained attention and have been demonstrated to perform well in SAR object detection tasks. Despite the good detection results achieved by existing CNN methods, they have overlooked the advantages of self-attention in modeling long-distance information.…”

“…Researchers have attempted to adapt and apply these methods to SAR target detection by making specific improvements. These improvements mainly include using stronger backbone networks [ 8 , 9 ], setting up multi-scale FPN layers [ 10 , 11 ], and designing loss functions more suitable for SAR tasks [ 12 , 13 ]. Meanwhile, real-time SAR target detection schemes [ 14 , 15 , 16 ] are also gradually developing, providing references for the practical application of SAR detection and recognition.…”

A Survey on Deep-Learning-Based Real-Time SAR Ship Detection

“…designed ShipDeNet-20, ShipDeNet-18, HyperLi-Net and DSDet respectively to detect ships in SAR images by training from scratch. We will review them in subsection H of V. Besides the above papers, Guo et al [35] and Peng et al [80] adopted CenterNet and YOLOX as the basic detector to train SAR ship detectors from scratch.…”

“…It achieved 97.4% AP50 on SSDD, 89.7% AP50 on HRSID with 18.38G FLOPs, 5.68M parameters and 7.01ms testing time on GeForceRTX2060 GPU. Peng et al [80] improved YOLOX with corner efficient intersection over union, adaptive-NMS, atrous convolution and coordinate attention mechanism for detecting sparse and small ships. It achieved 91.76% AP50 with 11ms testing time on HRSID RTX3060ti GPU.…”

IEC-YOLOX:A Lightweight Object Detection Algorithm for UAV Target Detection

“…The soft-merge algorithm incorporates the confidence scores and coordinate information of all prediction boxes to construct a fused prediction box that is closer to the real target box, thereby it mitigates issues related to loss detection and error detection. Hui Peng et al [25] proposed an Adaptive NMS method, which concatenates regression feature map and the previous feature map in the channel dimension. Then it employs convolutional operation to determine their distribution density.…”