SIA OpenIR  > 光电信息技术研究室
面向复杂背景下小目标检测的深度神经网络研究
Alternative TitleResearch on Deep Neural Network for Small Target Detection in Complex Background
鞠默然
Department光电信息技术研究室
Thesis Advisor罗海波
Keyword卷积神经网络 对地小目标检测 对空红外弱小目标检测 超分辨率重建 图像过滤
Pages131页
Degree Discipline模式识别与智能系统
Degree Name博士
2021-05-24
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract作为计算机视觉领域中的一个研究热点,目标检测已广泛应用于智能交通、智能监控、行为识别和医学图像辅助诊断等多个领域。其中,对地小目标检测和对空红外弱小目标检测可分别在对地作战和对空作战中用于侦察目标,有助于实现“先敌发现,先敌锁定和先敌命中”,对国防事业具有重要的应用价值。然而,对地小目标和对空红外弱小目标在图像中占有的像素非常少,特征不明显。因此,对地小目标检测和红外弱小目标检测更具有挑战性。随着计算机技术的高速发展,基于深度学习的目标检测算法成为目标检测领域中的主流方法。基于深度学习的目标检测算法可以将特征提取、特征融合、目标定位和分类等多个任务融合到一个网络中,实现对目标端到端的检测。综上所述,基于深度学习的对地小目标检测和对空红外弱小目标检测是一个具有显著应用价值且具有挑战性的研究课题。因此,本文分别对基于深度学习的对地小目标和对空红外弱小目标检测算法展开研究,本文的主要贡献如下:1. 分析并总结了影响小目标检测的关键因素。由于对地小目标和对空红外弱小目标包含像素少、特征不明显,易受周围环境的影响。因此,与常规目标相比,对地小目标和对空红外弱小目标检测更具有挑战性。为了设计适合于对地小目标和对空红外弱小目标检测的算法,本文首先对影响小目标检测的关键因素进行了分析和总结。并针对各个因素,提出相应的解决方法。2. 提出了面向小目标检测的图像预处理方法。在对地小目标检测任务中,输入图像的大部分区域都是背景,只有少部分区域是目标。根据此特点,提出了一种基于图像裁剪的对地小目标预处理方法。利用该方法,可以在不增加显存的条件下,实现网络对目标多尺度的训练,使网络变得更加鲁棒。另外,针对红外弱小目标信号“弱”、“小”且常常伴随着噪声干扰的特点,提出了基于局部对比度的红外弱小目标图像预处理方法。该方法可以利用简单的卷积操作和最大池化操作来实现,可以快速、有效地提高目标与背景之间的对比度。3. 提出了一种基于超分辨率重建的对地小目标检测网络SuperDet。作为影响小目标检测的一个关键因素,提高小目标的分辨率有助于提高检测网络对小目标的检测性能。SuperDet通过结合超分辨率重建技术和目标检测技术来提高对对地小目标的检测性能。SuperDet主要由两部分组成,分别是超分辨率模块和对地小目标检测模块。超分辨率模块实现了低分辨率图像到高分辨率图像的转换。将重建后的高分辨率图像作为输入送到对地小目标检测模块完成对对地小目标位置和类别的预测。为了选择合适的尺度来检测小目标,利用尺度匹配策略为对地小目标设计主干网络。随后提出了加权感受野融合模块,通过自适应融合具有不同感受野的特征来增加对地小目标周围的上下文信息。另外,针对对地小目标检测中正负样本数量不均衡的问题,提出了正样本选择策略。最后,分别采用GIOU loss、Focal loss和Binary Cross Entropy loss来对小目标的位置、置信度和类别进行回归。为了实现对SuperDet端到端的训练,采用多任务损失函数对超分辨率重建任务的损失和目标检测任务的损失进行回归。实验结果表明,SuperDet具有较为出色的对地小目标检测能力。 4. 提出了一种基于图像过滤的对空红外弱小目标检测网络FilterDet。FilterDet主要由两部分组成,分别是过滤模块和红外弱小目标检测模块。其中,过滤模块实现了对背景的抑制和对红外弱小目标的增强,可以过滤掉图像中的噪声和干扰。红外弱小目标检测模块将过滤后的图像作为输入,完成对红外弱小目标位置的预测。根据尺度匹配策略,采用沙漏结构来设计红外弱小目标检测模块的主干网络。另外,采用加权感受野融合模块来增加小目标周围的上下文信息。利用空间注意力机制来增强网络的表达能力,使红外弱小目标的特征响应得到加强。FilterDet将图像过滤任务和红外弱小目标检测任务融合到了一个网络中,实现了对红外弱小目标端到端的检测。实验结果表明,FilterDet具有较为出色的对空红外弱小目标检测性能。
Other AbstractAs a research hotspot in the field of computer vision, target detection has been widely used in intelligent transportation, intelligent monitoring, action recognition and medical image aided diagnosis. Ground small target detection and aerial infrared small target detection can be used to perform target reconnaissance in to-air operations and to-ground operations respectively, which is helpful to realize “first enemy detection, first enemy lock and first enemy hit”, and has important application value to national defense. Compared with generic targets, ground small targets and aerial infrared dim and small targets cover only a small part of an image and the features of these targets are inconspicuous. Therefore, ground small target detection and aerial infrared dim and small target detection are more challenging. With the rapid development of computer technology, target detection algorithm based on deep learning has become the mainstream method in the field of target detection. That is because the target detection algorithms based on deep learning fuse multiple tasks such as feature extraction, feature fusion, target location and classification into a network, which can perform target detection in an end-to-end way. To sum up, ground small target detection and aerial infrared dim and small target detection based on deep learning is a research subject with significant application value and challenges. Therefore, this dissertation researches on the ground small target detection algorithm based on deep learning and aerial infrared dim and small target detection algorithm based on deep learning respectively. The main contributions of this dissertation are as follows: 1. This dissertation analyzes and summarizes the factors affecting the detection performance of the small targets. Ground small targets and aerial infrared dim and small targets cover few pixels in an image and their features are not obvious. In addition, they are easily affected by the surrounding environment. Therefore, compared with the general target detection, ground small target detection and aerial infrared dim and small target detection are more challenging. In order to design algorithms suitable for ground small target detection and aerial infrared dim and small target detection, this dissertation firstly analyzes and summarizes the key factors affecting small target detection and then puts foward solutions in view of each factor. 2. Image preprocessing methods for small target detection are proposed. In ground small target detection task, most regions of the input image are background and the targets only cover a few regions. According to this characteristic, an image cropping method for ground small target image preprocessing is introduced. By this method, the network can achieve multi-scale training on the same target without the increasement of memory, which is helpful to make the network more robust. In addition, aiming at the characteristics that the signal of infrared dim and small targets is dim and the targets are often accompanied by noise and interference, a local contrast method for infrared dim and small target image preprocessing is introduced. This image preprocessing method can be achieved by using simple convolution and max-pooling operation, which can enhance the contrast between the targets and background efficiently. 3. Aiming at ground small target detection, a ground small target detection network based on super-resolution reconstruction, called SuperDet, is proposed. As a key factor affecting the detection performance of small targets, the resolution of target plays an important role. Improving the resolution of small targets is conducive to improving the detection performance of the network. Therefore, a ground small target detection network SuperDet is proposed by combining the super-resolution reconstruction technology with the target detection algorithm. SuperDet consists of two parts, namely, super-resolution reconstruction module and target detection module. The super-resolution reconstruction module aims to recover a high resolution image from its low resolution counter part. The target detection module takes the reconstructed high resolution image as input and predict the location and the category of the ground small targets. In order to select appropriate scales for ground small target detection, the scale matching strategy is proposed. According to the scale matching strategy, we design the backbone for the ground small targets. In addition, a weighted receptive field fusion module is introduced to increase the context information around the targets by merging the features with different receptive field. What is more, a positive sample selection strategy is proposed to alleviate the imbalance between the positive samples and negative ones in ground small target detection. Finally, GIOU Loss, Focal Loss and Binary Cross Entropy Loss are used to regress the position, confidence and category of the targets. To perform the end-to-end training of SuperDet, multi-task loss function is used to regress the loss of super-resolution reconstruction task and ground small target detection. The experimental results show that SuperDet has excellent detection ability for ground small targets. 4. Aiming at aerial infrared dim and small target detection, an infrared dim and small target detection network based on image filtering, called FilterDet, is proposed. FilterDet consists of two parts, namely image filtering module and infrared dim and small target detection module. The image filtering module can filter out the noise and enhance the targets. The infrared dim and small target detection module takes the filtered image as input and predicts the location of the infrared dim and small targets. According to the scale matching strategy, the hourglass structure is used to design the backbone of infrared dim and small target detection module. In addition, the weighted receptive field fusion module is used to increase the context information around the small targets. Spatial attention mechanism is applied to improve the expressive ability of the network, which enhances the response of infrared dim and small targets. FilterDet integrates the image filtering task and the infared dim and small target detection task into one network, which realizes the end-to-end detection of the infrared dim and small targets. The experimental results show that FilterDet has excellent detection performance for aerial infrared dim and small targets.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/29010
Collection光电信息技术研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
鞠默然. 面向复杂背景下小目标检测的深度神经网络研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
面向复杂背景下小目标检测的深度神经网络研(21170KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[鞠默然]'s Articles
Baidu academic
Similar articles in Baidu academic
[鞠默然]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[鞠默然]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.