基于多尺度上下文提取的小样本野生动物检测  

Few-shot wildlife detection based on multi-scale context extraction

作　　者：刘珂 林珊玲 师欣雨 林坚普 吕珊红 林志贤[1,2] 郭太良 LIU Ke;LIN Shanling;SHI Xinyu;LIN Jianpu;LÜShanhong;LIN Zhixian;GUO Tailiang(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362251,China;Fujian Science and Technology Innovation Laboratory for Photoelectric Information,Fuzhou 350116,China)

机构地区：[1]福州大学先进制造学院,福建泉州362251 [2]中国福建光电信息科学与技术实验室,福建福州350116

出　　处：《液晶与显示》2025年第3期516-526,共11页Chinese Journal of Liquid Crystals and Displays

基　　金：国家重点研发计划(No.2021YFB3600603)。

摘　　要：针对野生动物数据集样本量小、目标尺度多变所导致的野生动物检测困难以及检测精度低等问题,提出一种基于多尺度上下文提取的小样本野生动物检测(MS-FSWD)算法。首先,通过多尺度上下文提取模块增强模型对不同尺度的野生动物的感知能力,提高检测性能;其次,引入Res2Net作为原型校准模块的强分类网络对分类器输出的分类分数进行校正;然后,在RPN中加入置换注意力机制,增强目标区域的特征图,弱化背景信息;最后,将平衡L1损失作为定位损失函数,提升目标定位性能。实验结果表明,相比DeFRCN算法,MS-FSWD在小样本野生动物数据集FSWA上,1-shot和3-shot检测任务中新类AP50分别提升了9.9%和6.6%;在公共数据集PASCAL VOC上,MS-FSWD最高提升了12.6%。与VFA算法相比,在PASCAL VOC数据集Novel Set 3的10-shot任务中,新类AP50提升了3.3%。In order to solve the problems of difficulties and low detection accuracy caused by the small sample size and variable target scale of wildlife datasets,a few-shot wildlife detection(MS-FSWD)algorithm based on multi-scale context extraction was proposed.Firstly,the multi-scale context extraction module was used to enhance the perception ability of the model for wildlife at different scales and improve the detection performance.Secondly,Res2Net was introduced as a strong classification network of the prototype calibration module to correct the class scores output by the classifier.Then,the shuffle attention mechanism was added to the RPN to enhance the feature map of the target region and weaken the background information.Finally,using the Balanced L1 Loss as the localization loss function improves the target positioning performance.Experimental results show that compared with the DeFRCN algorithm,MS-FSWD improves the novel class AP50 by 9.9%and 6.6%respectively in the 1-shot and 3-shot detection tasks on the few-shot wildlife dataset FSWA.On the public dataset PASCAL VOC,MS-FSWD is increased by up to 12.6%.Compared with the VFA algorithm,in the 10-shot task of the PASCAL VOC dataset Novel Set 3,the novel class AP50 is increased by 3.3%.

关 键 词：小样本目标检测 野生动物检测 迁移学习 多尺度上下文提取 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]