基于深度学习的船体附着物识别方法

Recognition Methods of Hull Fouling Based on Deep Learning

船体附着物对船舶航行有重要影响。为实现对附着物清洗时的自动快速识别,首先构建1个种类较完善的附着物图像数据集,并针对真实数据样本数量有限且较难获取导致目标识别的准确率较低的问题,提出一种基于深度卷积生成式对抗网络(DCGAN)和改进的Efficient Net的船体附着物识别方法。该方法通过DCGAN对样本数据进行增强,分别采用Google Net、VGG16、Mobile Net、Reg Net和Vision transformer这5个训练模型,以微调的方式对样本进行训练识别;同时,基于DCGAN增强的数据集对Efficient Net进行改进并训练,采用Fused-MBConv模块替换原来网络结构中的MBConv模块,简化浅层网络结构。结果表明:基于DCGAN数据增强的方法有助于提高模型的识别准确率;改进的Efficient Net模型的识别准确率提升了4.9%,并且训练时间也有所缩短。提出的方法能满足船体附着物的快速识别需求。

Hull fouling has an important impact on ship navigation. In order to realize the automatic and rapid recognition of hull fouling during cleaning, a complete set of hull fouling image datasets is constructed. Aiming at the low accuracy of target recognition due to insufficient samples of hull fouling and difficult to obtain, a target recognition of hull fouling method based on(DCGAN) and improved EfficientNet network is proposed. The method firstly applies DCGAN to enhance the sample data of hull fouling. Then, the five training models of Google Net, VGG16, MobileNet, RegNet and Vision transformer are used to train and recognize the samples in a fine-tuned manner. Finally, to improve the EfficientNet network, the Fused-MBConv module is used to replace the MBConv module of the original structure, which can simplify the module structure of the shallow network.The results show that the DCGAN data augmentation method can improve the recognition accuracy of the hull fouling. The recognition accuracy of the improved Efficient Net model is increased by 4.9% on the image datasets of hull fouling, and the training time also reduces.The proposed method can meet the needs of rapid recognition of hull fouling. The proposed method meets the requirements of rapid identification of hull attachments.