基于超分辨率重建的小目标智能检测算法  

Intelligent Detection Algorithm for Small Targets Based on Super-Resolution Reconstruction

作　　者：蔡心悦 周杨[1,2,3] 胡校飞 吕亮 赵璐颖[1,4] 彭杨钊 Cai Xinyue;Zhou Yang;Hu Xiaofei;LüLiang;Zhao Luying;Peng Yangzhao(Institute of Geospatial Information,Information Engineer University,Zhengzhou 450001,Henan,China;Collaborative Innovation Center of Geo-Information Technology for Smart Central Plains,Henan Province,Zhengzhou 450001,Henan,China;Key Laboratory of Spatiotemporal Perception and Intelligent Processing,Ministry of Natural Resources,Zhengzhou 450001,Henan,China;Henan Technical College of Construction,Zhengzhou 450001,Henan,China)

机构地区：[1]信息工程大学地理空间信息学院,河南郑州450001 [2]智慧中原地理信息技术河南省协同创新中心,河南郑州450001 [3]时空感知与智能处理自然资源部重点实验室,河南郑州450001 [4]河南建筑职业技术学院,河南郑州450001

出　　处：《激光与光电子学进展》2023年第12期41-49,共9页Laser & Optoelectronics Progress

摘　　要：针对小目标占有像素少导致检测精确率低的问题,提出一种基于超分辨率重建的小目标检测算法。首先,通过图像预处理对高分辨率图像分块并筛选出含有目标的子图像;其次,构建超分辨率锐化增强模块,引入锐化图像和锐化损失,以获得边缘更清晰的高分辨率子图像;然后,采用多尺度锐化目标检测模块检测目标,该模块添加边缘锐化模型,在深层特征层中进一步锐化图像边缘,弥补深层卷积对细节的损失;最后,根据子图像编号将小目标检测结果回归到原图像中,完成小目标图像检测。在PASCAL VOC数据集和COCO 2017数据集上的实验结果表明,所提算法的平均精确率(mAP)分别为85.3%和54.0%,对COCO数据集的小目标检测精确率为43.5%,高于次优值9.7个百分点。因此,所提算法可以有效减少小目标漏检的次数,提高检测精确率。A small target detection algorithm based on super-resolution reconstruction is proposed to solve the problem of low detection accuracy of small targets occupying a few pixels.First,a high-resolution image is segmented via image preprocessing and sub-images containing targets are filtered out.Second,a super-resolution sharpening enhancement module is constructed,and the sharpening image and sharpening loss are introduced to obtain high-resolution sub-images with clearer edges.Subsequently,a multi-scale sharpening target detection module is used to detect the target;it uses an edge-sharpening model to further enhance the image edges of the deep feature layer to compensate for the loss in details due to deep convolution.Finally,the small-target detection results are returned in the original image based on the sub-image number used to complete small target image detection.The proposed detection algorithm is then verified using the PASCAL VOC and COCO 2017 datasets,where the average accuracies(mAP)are 85.3%and 54.0%,respectively.Moreover,the small target detection accuracy of the COCO dataset is 43.5%,which is 9.7 percentage points higher than the suboptimal value.Therefore,the proposed algorithm can effectively reduce the number of times small targets are missed during detection,thus improving the detection accuracy.

关 键 词：图像处理 小目标检测 超分辨率增强 卷积神经网络 多尺度特征融合 边缘锐化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]