文献类型：期刊文献

中文题名：基于增强特征提取的森林遥感图像行人小目标检测网络

英文题名：Enhanced feature extraction network for small pedestrian detection in forest remote sensing images

作者：李春燕[1] 王超[1] 金星[1] 符利勇[2] 业巧林[1,3]

第一作者：李春燕

机构：[1]南京林业大学信息科学技术学院、人工智能学院,南京210037;[2]中国林业科学研究院资源信息研究所,北京100091;[3]南京晓庄学院智能信息处理重点实验室,南京211171

年份：2024

卷号：9

期号：4

起止页码：130-139

中文期刊名：林业工程学报

外文期刊名：Journal of Forestry Engineering

收录：CSTPCD;;Scopus;北大核心:【北大核心2023】；CSCD:【CSCD_E2023_2024】；

基金：国家自然科学基金(62072246)。

语种：中文

中文关键词：遥感图像;行人小目标;增强特征提取;感受野增强模块;GKCLOSS损失函数

外文关键词：remote sensing image;pedestrian mini-target;enhanced feature extraction;receptive field block attention module;GKCLOSS loss function

分类号：TP319.4

摘要：林业作业人员常常由于复杂地貌而面对坠落、森林火灾等威胁,卷积神经网络结合无人机巡查的方法已经成为主流防范、搜救措施,但遥感图片中的行人小目标有特征少、定位精度要求高以及极易受背景信息干扰的特点。为了能够使森林遥感图片行人小目标检测的精度达到预期,在YOLOv4方法的基础上针对上述特点设计了增强特征提取的目标检测网络(EFEN),通过构建感受野增强模块(RFBA)并结合CBAM注意力机制,在充分利用遥感图片中的丰富上下文信息之余,对相关信息进行动态选择,增强特征的表示能力;基于高斯分布思想,将归一化Wasserstein距离与CIOU结合,提出了一种新的损失函数(GKCLOSS),降低了小目标检测任务中对位置偏差的敏感性;引入一种自适应分割训练检测策略,平衡正负样本,提高目标检测的准确性,进一步提高了检测精度。以河北省张家口市崇礼区采集的无人机行人图像为研究对象,实验表明,EFEN框架在小目标检测方面优于现有的深度学习网络,在与SSD、YOLOv5、YOLOv7等算法比较中平均查准率(mAP)均有所提升,在上述数据集上,mAP高达39.10%,证明了此方法对行人小目标数据的有效性。
The key of mitigating property and casualty losses caused by forest disasters lies in accurately locating individuals involved.However,in the task of unmanned aerial vehicle(UAV)remote sensing image detection,pedestrians'detection in such complex scenes,as small objects,are prone to missing detection and deviation due to information distortion and loss.To address the challenge of accurately detecting pedestrians in complex forest scenes,a new object detection network with enhanced feature extraction(EFEN)was proposed in this study,which incorporated the receptive field enhancement module,and optimized the loss function and the data preprocessing.The receptive field block(RFB)was reconstructed,being called RFBA,which was embedded into the YOLOv4 backbone for the expansion of the network receptive field.RFBA eliminated the risk of information loss in dilated convolution,while retaining the high performance of the original module in processing multi-scale and contextual features.Howe-ver,the training of model was inevitably disturbed by uninformative pixels,which were implicated by the enlargement of receptive field.Thus,convolutional block attention module(CBAM)was integrated into the network,which enhanced the processing ability of the network when dealing with information of various scales,shapes,and directions.CBAM quantified the information features convey by analyzing the correlation between features across multiple channels and different spatial locations and assigns weights to filters out some features with less information and contribution.The loss function was further optimized by combining the Normalized Wasserstein Distance(NWD)similarity measurement with CIOU,being called GKCLOSS(Gaussian Kullback-Leibler and Complete-IOU loss),which alleviated small objects'sensitivity of positioning offset.Besides,a new segmentation training strategy was involved to deal with the problem of imbalanced samples of small targets in remote sensing images.The image was segmented and the recognition areas were adaptively screened,which contained or were around the target.Extensive experiments were conducted on the Chongli Winter Olympic Games pedestrian dataset,showcasing the remarkable performance of the EFEN framework in small object detection,with the mean average precision(mAP)on the above dataset up to 39.10%,and the mAP was improved by EFEN,compared to SSD,YOLOv5 and YOLOv7 algorithms,which de-monstrated the effectiveness of this method during small object detection task.