改进YOLOX的SAR近岸区域船只检测方法

Improved YOLOX SAR Near-Shore Area Ship Detection Method

针对SAR近岸区域船只检测准确率低与虚警率高的问题,提出一种基于改进注意力机制与旋转框的SAR近岸区域船只检测方法。该方法首先通过改进坐标注意力机制并引入至特征提取网络中,提升网络的特征提取能力;其次,增加角度分类头,并引入二维高斯分布,计算预测分布与目标分布的KL散度评估旋转框损失值,完成目标的角度信息提取;再基于YOLOX算法中的无锚框(AF)机制,减少候选框冗余,使模型轻量化并进一步提高定位精度。最后在公开数据集offical-ssdd上进行测试,在嵌入式平台(NVIDIA Jetson AGX Xavier)上对模型进行推理验证。该算法模型计算参数仅1.14 M,在近岸情况下平均检测精度较YOLOX模型提高了18.77%,总体检测精度达到94.2%。验证结果表明,该算法适用于复杂场景下任意方向的密集船只目标检测,满足实时处理需求。

To solve the problem of low accuracy and high false alarm rate of synthetic aperture radar(SAR)nearshore area vessel detection,a new SAR nearshore area vessel detection method based on improved attention mechanism and rotating frame is proposed.Firstly,the feature extraction capability of the network was enhanced by improving the coordinate attention mechanism and introducing it into the feature extraction network.Secondly,the angle classification head was added and the two-dimensional Gaussian distribution was introduced to calculate the KL divergence between the prediction distribution and the target distribution,so as to evaluate the loss value of the rotating frame and complete the angle information extraction of the target.Then,based on the anchor frameless(AF)mechanism of YOLOX algorithm,the model can be made lightweight and the positioning accuracy can be further improved by reducing the redundancy of candidate frames.Finally,the model was tested on the open dataset Offical-SSDD,and the inference verification was performed on the embedded platform(NVIDIA Jetson AGX Xavier).The calculation parameter of the algorithm model is only 1.14M,and the average detection accuracy of the algorithm model is 18.77%,higher than that of the YOLOX model in the nearshore condition,and the overall detection accuracy reaches 94.2%.The verification results show that the algorithm is suitable for dense ship target detection in any direction in complex scenes and can meet the requirements of real-time processing.