摘要
近年来,随着舰船目标检测算法的不断创新,舰船检测性能有了显著的改善.然而,由于算法模型复杂度高、规模大,部署在资源有限的环境下存在一定的挑战.针对上述问题提出了一种新的轻量化模型GFPN-YOLOX,首先,在主干提取网络中引入感受野注意力卷积,让模型可以提取到多尺度目标的特征信息,从而增强模型的特征处理能力;其次,设计一种新的特征金字塔融合模块,用于增强遥感数据的语义信息与低维的位置信息之间的信息交流;最后,在特征金字塔融合模块和检测头部中引入了深度可分离卷积,用于降低模型的复杂度和存储需求.仿真结果表明,与原模型YOLOX-Tiny算法相比,模型参数量减少了 29%,计算量降低了 38%,但精度仅有 0.4%的损失.通过与最新算法对比,验证了该模型在遥感数据检测场景中的有效性.
Abstract
In recent years,there has been significant improvement in ship detection performance due to continuous innovation in ship target detection algorithms.However,deploying these algorithms in resource-limited environments poses certain challenges because of their high complexity and large scale.To address these issues,this paper proposes a new lightweight model GFPN-YOLOX.Firstly,the introduction of receptor field attention convolution into the backbone extraction network enhances the feature processing capability by extracting feature information from multi-scale targets.Secondly,a new feature pyramid fusion module is designed to improve information exchange between semantic information of remote sensing data and low-dimensional location information.Finally,deep separable convolution is introduced to reduce the complexity and storage requirements of the model.Compared to the original YOLOX-Tiny algorithm,this model reduces parameter count by 29%and calculation amount by 38%,while only losing accuracy by 0.4%.Its effectiveness in remote sensing data detection scenarios is verified compared to the latest algorithm.
维普
万方数据