改进YOLOX在近岸船舶检测中的应用

Application of Improved YOLOX in Inshore Ship Inspection

为了解决近岸船舶检测时目标尺度变化大,环境干扰严重等问题,提出了一种改进YOLOX的无锚框检测算法。首先,在主干网络中引入CoT模块,通过动态利用上下文信息来增强表达能力,降低环境干扰的影响;其次,将SimAM注意力嵌在特征金字塔和检测头之间,丰富语义信息,提升小目标检测精度。再利用CIOU来取代原有损失函数,以提高收敛速度;最后,使用深度可分离卷积替换特征金字塔中普通卷积,减少参数量,提升检测速度。实验结果表明:在SeaShips数据集上,改进后模型在减少参数量的同时,精度提高了6.73%,均值平均精度(mAP)达到了96.63%,检测速度达到了48.6帧/s,能够实时、高精度地检测近岸船舶。

To solve the problems of large changes in target scale and serious environmental interference during nearshore ship detection,an improved anchorless frame detection algorithm of YOLOX is proposed.Firstly,the contextual transformer(CoT)module is introduced in the backbone network to enhance the expression capability and improve the problem of severe environmental interference by dynamically using contextual information.Secondly,SimAM attention is embedded between the feature pyramid and detection head to enrich semantic information and improve small target detection accuracy.Furtherly,CIOU is used to replace the original loss function to improve the convergence speed.Finally,the depth-separable convolution is used to replace the ordinary convolution in the feature pyramid to reduce the number of parameters and improve the detection speed.The experimental results show that on the SeaShips dataset,the improved model improves the accuracy by 6.73%,mAP reaches 96.63%,and detection speed reaches 48.6 frames per second while reducing the number of parameters,which can detect near-shore ships in real time and with high accuracy.