基于注意力机制的海上小目标重识别方法

Re-identification Method of Small Target in Ocean Environment Based on Attention Mechanism

在海上由于涉及到目标的遮挡、背景及照明等的巨大变化,小目标重识别是一项极具挑战性的计算机视觉任务。针对传统小目标识别算法泛化性差,在目标重识别中特征提取不够充分有效的问题,提出一种基于通道和空间注意力机制的目标重识别方法。首先,通过将CBAM机制嵌入到重识别模型的骨干网中,通过引入通道注意力机制和空间注意力机制,反馈了更加关键的特征信息。针对骨干输出的三维特征量,为得到更有效的信息,沿垂直进行不同比例的平均分块,从而在多粒度的情况下,从全局和局部更加全面地关注图像信息,同时利用多损失函数分别优化模型,提升模型的可区分性。最后将梯度中心化算法引入Adam优化器,提升网络模型的训练速度和泛化能力,为海上小目标重识别提供了一种新的研究思路与方法。

In ocean environment due to the huge changes involved in the occlusion,background and lighting,object re-identification is a very challenging computer vision task.Traditional small target recognition algorithms has poor generalization ability,aiming at the problem of insufficient feature extraction in object re-recognition,a object re-identification method was proposed based on channel and spatial attention mechanism.First of all,the CBAM mechanism was embeded into the backbone network of the re-identification model,which introduced the channel attention mechanism and the spatial attention mechanism to feed back more critical feature information.For the three-dimensional feature output of the backbone,in order to obtain more effective information,average partitioning of different proportions along the vertical was performed,so that in the case of multiple-granularity features,more attention could be paid to the image information from the global and local aspects,and multiple loss functions were used.The models were optimized separately to improve the distinguish-ability of the models.Finally,the gradient centralization algorithm was introduced into the Adam optimizer to improve the training speed and generalization ability of the network model.It provides a new research idea and method for target re-identification in ocean.