基于层间分类一致性准则的舰船目标细粒度识别方法

Fine-grained ship detection based on consistency criteria of hierarchical classification

舰船目标的层次化、细粒度识别在军事和民用领域均有重要意义。现有细粒度识别方法一般需要部件级精细标注或采用注意力机制提取关键特征,但并未有效利用舰船目标层次化分类体系中本身所蕴含的隶属关系信息提高细粒度识别精度。针对舰船目标的层次化分类问题,建立了舰船目标多层级一致性分类数学模型,提出了一种基于层间强一致性分类准则的细粒度识别方法,设计了层间一致性分类损失函数,并构建了多层级兼容舰船目标细粒度识别网络(MLCDet)。经试验验证,该方法有效、鲁棒,资源开销小,能够有效利用分类体系中各类别间的隶属关系提升目标识别精度。在无需部件级标注信息的前提下,将mAP提高了1.3%,与此同时,模型总参数量仅增加0.02%,推断速度不变。

Hierarchical and fine-grained detection of ships is essential in both military and civilian applications.Existing fine-grained detection approaches often need part-level labeling or an attention mechanism to retrieve key features.However,they do not properly exploit the affiliation information inherent in the hierarchical categorization structure of ships to increase fine-grained detection performance.Aiming at ships'hierarchical classification,we built a multi-level consistent classification mathematical model for ships.This paper proposed a fine-grained detection method and loss function based on the strict consistency criterion across multiple classification levels and created a multi-level compatible fine-grained ship detection network(MLCDet).The experimental results show that the method is effective,robust,has low resource consumption,and can effectively utilize the affiliation between categories in the classification system to improve object detection accuracy.Without the requirement for parts annotation information,mAP is increased by 1.3 percent.At the same time,the total model parameters are only increased by 0.02 percent,while the inference speed remains unchanged.