CN114842324A - Method and system for detecting disguised target based on learning neural network - Google Patents
Method and system for detecting disguised target based on learning neural networkDownload PDFInfo
- Publication number
- CN114842324A CN114842324ACN202210257144.XACN202210257144ACN114842324ACN 114842324 ACN114842324 ACN 114842324ACN 202210257144 ACN202210257144 ACN 202210257144ACN 114842324 ACN114842324 ACN 114842324A
- Authority
- CN
- China
- Prior art keywords
- camouflage
- target
- representation
- recognition
- level
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
Description
Translated fromChinese技术领域technical field
本发明属于伪装目标检测领域，具体涉及一种基于学习神经网络的伪装目标检测方法及系统。The invention belongs to the field of camouflage target detection, in particular to a camouflage target detection method and system based on a learning neural network.
背景技术Background technique
伪装目标识别是许多应用的关键技术，包括计算机视觉领域(可用于搜索和救援工作，或寻找稀有物种)、医学图像分析(如息肉分割和肺炎分割)、农业领域(如蝗虫入侵监控)和艺术领域(用于真实感图像融合或艺术消遣)。一般来说，伪装目标识别仍然是一个具有挑战性的问题，主要原因有两点，一是伪装物体与其背景之间具有高度相似性，伪装目标表征难以提取；二是伪装目标的上下文细节信息难以有效增强。自然界中有很多生物已经进化到与周围环境相混淆，甚至更难被检测。Camouflaged object recognition is a key technique for many applications, including the field of computer vision (which can be used in search and rescue efforts, or to find rare species), medical image analysis (such as polyp segmentation and pneumonia segmentation), agricultural fields (such as locust invasion surveillance), and the arts Field (for photorealistic image fusion or artistic recreation). Generally speaking, camouflage target recognition is still a challenging problem. There are two main reasons. One is that the camouflage object is highly similar to its background, so it is difficult to extract the camouflage target representation; the other is that the context details of the camouflage target are difficult to extract Effective enhancement. There are many organisms in nature that have evolved to be confused with their surroundings and even harder to detect.
在伪装目标检测中，由于环境与目标的高相似性，导致提取得到的目标表征并不精确，并且上下文的细节信息没有被深度挖掘出来，并且伪装目标的检测准确性也高度依赖于阈值，现有技术中目标检测方法并不能达到良好的性能，如何选择出合适的阈值本身也是一个困难的问题。因此，伪装目标识别研究具有十分重要的意义，近年来引起了广泛的关注。In the detection of camouflaged targets, due to the high similarity between the environment and the target, the extracted target representation is not accurate, and the details of the context are not deeply excavated, and the detection accuracy of the camouflaged target is also highly dependent on the threshold. The target detection methods in the prior art cannot achieve good performance, and how to choose an appropriate threshold is also a difficult problem. Therefore, the research on camouflage target recognition is of great significance and has attracted extensive attention in recent years.
发明内容SUMMARY OF THE INVENTION
本发明的目的在于提供一种基于学习神经网络的伪装目标检测方法及系统，将伪装目标表征通过递进分片，的层级方式输入学习神经网络中，提高伪装目标的检测准确性。The purpose of the present invention is to provide a camouflage target detection method and system based on a learning neural network. The camouflage target representation is input into the learning neural network in a hierarchical manner through progressive slicing, so as to improve the detection accuracy of the camouflage target.
为达到上述目的，本发明所采用的技术方案是：In order to achieve the above object, the technical scheme adopted in the present invention is:
本发明第一方面提供了一种基于学习神经网络的伪装目标检测方法，包括：A first aspect of the present invention provides a method for detecting camouflaged targets based on a learning neural network, including:
采集检测图像输入至预先训练好的目标搜索模型，获得低级到高级依次设置的多层伪装目标表征；Collect detection images and input them to the pre-trained target search model to obtain multi-layer camouflage target representations set in sequence from low-level to high-level;
由各层伪装目标表征中提取伪装表征分片，并输入至预先训练好的目标识别模型获得相应等级的多个伪装识别特征；Extract camouflage representation fragments from each layer of camouflage target representation, and input them into the pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels;
将各伪装识别特征进行卷积获得伪装识别层级图像，将各伪装识别层级图像进行叠加获得伪装识别图像；Convolving each camouflage recognition feature to obtain a camouflage recognition level image, and superimposing each camouflage recognition level image to obtain a camouflage recognition image;
所述目标搜索模型和目标识别模型的训练过程包括：The training process of the target search model and target recognition model includes:
采集包含伪装目标的训练图像，获取训练数据集；Collect training images containing camouflaged targets to obtain training data sets;
将通过训练数据集对目标搜索模型和目标识别模型进行训练，计算训练过程中伪装识别的结构化损失值S，根据结构化损失值S迭代调整目标搜索模型和目标识别模型的参数，获得设定要求的目标搜索模型和目标识别模型。The target search model and target recognition model will be trained through the training data set, the structural loss value S of camouflage recognition during the training process will be calculated, and the parameters of the target search model and target recognition model will be iteratively adjusted according to the structural loss value S, and the set value will be obtained. The required target search model and target recognition model.
优选的，所述采集检测图像输入至预先训练好的目标搜索模型，获得低级到高级依次设置的多层伪装目标表征的方法包括：Preferably, the collection and detection images are input to the pre-trained target search model, and the method for obtaining the multi-layer camouflage target representations set sequentially from low-level to high-level includes:
将检测图像输入至目标搜索模型中的Res2Net50残差网络，输出低级到高级依次设置的四层伪装表征f；Input the detection image to the Res2Net50 residual network in the target search model, and output the four-layer camouflage representation f that is set in sequence from low-level to high-level;
通过感觉野网络对各层伪装表征f的局部细节信息放大，分别获得伪装目标表征fr2、伪装目标表征fr3、伪装目标表征fr4和伪装目标表征fr5。Through the sensory field network, the local detail information of each layer of camouflage representation f is amplified, and the camouflage target representation fr2 , the camouflage target representation fr3 , the camouflage target representation fr4 and the camouflage target representation fr5 are obtained respectively.
优选的，由各层伪装目标表征中提取伪装表征分片的方法包括：Preferably, the method for extracting camouflage representation fragments from each layer of camouflage target representations includes:
提取伪装目标表征fr5的四分片作为伪装表征分片fr′5；提取伪装目标表征fr4的二分片作为伪装表征分片fr′4；提取伪装目标表征fr3的整体作为伪装表征分片fr′3；提取伪装目标表征fr2的整体作为伪装表征分片fr′2。extracting the quadrant of the camouflage target representation fr5 as the camouflage representation segment fr′5 ; extracting the second segment of the camouflage target representation fr4 as the camouflage representation segment fr′4 ; extracting the whole of the camouflage target representation fr3 as the camouflage representation segment fr′3 ; extract the whole of the camouflage target representation fr2 as the camouflage representation fragment fr′2 .
优选的，伪装表征分片输入至预先训练好的目标识别模型获得相应等级的多个伪装识别特征的方法包括：Preferably, the method for obtaining multiple camouflage recognition features of corresponding levels by inputting the camouflage representation into slices into a pre-trained target recognition model includes:
按照由低级到高级的顺序将伪装表征分片fr′5、伪装表征分片fr′4、伪装表征分片fr′3和伪装表征分片fr′2依次输入至目标识别模型；Input the camouflage representation segment fr'5 , the camouflage representation segment fr'4 , the camouflage representation segment fr'3 and the camouflage representation segment fr'2 into the target recognition model in order from low level to high level;
对最低级的伪装表征分片fr′5进行小波注意力变换提取伪装识别特征r5；Perform wavelet attention transform on the lowest level camouflage representation fragment fr'5 to extract camouflage identification feature r5 ;
将下一级伪装表征分片rfn和上一级提取的伪装识别特征rn+1结合后，通过小波注意力变换依次获得伪装识别特征rn。After combining the next-level camouflage representation fragment rfn and the camouflage recognition feature rn+1 extracted by the previous level, the camouflage recognition feature rn is sequentially obtained through wavelet attention transformation.
优选的，将下一级伪装表征分片rfn和上一级提取的伪装识别特征rn+1结合后，通过小波注意力变换依次获得伪装识别特征的方法包括：Preferably, after combining the next-level camouflage representation fragment rfn with the camouflage identification features rn+1 extracted at the previous level, the method for sequentially obtaining the camouflage identification features through wavelet attention transformation includes:
将下一级伪装表征分片rfn和上一级提取的伪装识别特征rn+1进行结合后进行卷积，获得低频分量LL、高频分量LH、高频分量HL和高频分量HH；Combining the next-level camouflage representation fragment rfn and the camouflage identification feature rn+1 extracted by the previous level and convolving to obtain a low-frequency component LL, a high-frequency component LH, a high-frequency component HL, and a high-frequency component HH;
由所述低频分量LL捕获伪装目标的伪装分割信息和轮廓信息；通过注意力机制AM和含有实例归一化残差块ResIN对低频分量LL进行加强，叠加伪装分割信息和轮廓信息，获得低频分量伪装特征rLL；The camouflage segmentation information and contour information of the camouflaged target are captured by the low-frequency component LL; the low-frequency component LL is enhanced through the attention mechanism AM and the residual block ResIN containing instance normalization, and the camouflage segmentation information and contour information are superimposed to obtain the low-frequency component. camouflage feature rLL ;
将高频分量LH、高频分量HL和高频分量HH分别通过归一化残差块ResB和注意力机制AM进行增强，获得高频分量伪装特征rLH、高频分量伪装特征rHL和高频分量伪装特征rHH；The high-frequency component LH, high-frequency component HL and high-frequency component HH are respectively enhanced by the normalized residual block ResB and the attention mechanism AM to obtain the high-frequency component camouflage feature rLH , the high-frequency component camouflage feature rHL and the high-frequency component camouflage feature r HL . frequency component camouflage feature rHH ;
将低频分量伪装特征rLL、高频分量伪装特征rLH、高频分量伪装特征rHL和高频分量伪装特征rHH进行转置卷积TC，获得伪装识别特征rk。The low-frequency component camouflage feature rLL , the high-frequency component camouflage feature rLH , the high-frequency component camouflage feature rHL and the high-frequency component camouflage feature rHH are transposed convolution TC to obtain the camouflage identification featurerk .
优选的，计算训练过程中伪装识别的结构化损失值S，计算公式为：Preferably, the structural loss value S of camouflage recognition in the training process is calculated, and the calculation formula is:
S＝α*So+(1-α)*SrS=α*So +(1-α)*Sr
So＝μ*OFG+(1-μ)*OBGSo = μ*OFG +(1-μ)*OBG
其中，So表示为面向物体的结构相似性度量；Sr表示为面向区域的结构相似性度量；α表示为权重因子；Wk表示为第k个伪装表征分片的权值；K表示为伪装表征分片的总量；k表示伪装表征分片的序列号；μ为前景面积所占图像的比例；OBG表示为背景度量；OFG表示为前景度量。Among them, So is the object-oriented structural similarity measure; Sr is the region-oriented structural similarity measure; α is the weight factor; Wk is the weight of the k-th camouflage representation slice; K is expressed as The total number of camouflage representation fragments; k represents the serial number of camouflage representation fragments; μ is the proportion of the foreground area in the image; OBG represents the background metric; OFG represents the foreground metric.
优选的，计算所述背景度量OBG和前景度量OFG，表达公式为：Preferably, the background metric OBG and the foreground metric OFG are calculated, and the expression formula is:
其中，DFG表示为伪装预测图与伪装表征分片的前景区域差别程度，，xFG表示伪装目标预测图的前景区域的概率值，yFG分别表示伪装表征分片的前景区域的概率值，
本发明第二方面提供了一种基于学习神经网络的伪装目标检测系统，包括：A second aspect of the present invention provides a camouflaged target detection system based on a learning neural network, including:
采集模块，采集检测图像；The acquisition module collects inspection images;
目标搜索模块，用于将采集的检测图像输入至预先训练好的目标搜索模型，获得低级到高级依次设置的多层伪装目标表征；The target search module is used to input the collected detection images into the pre-trained target search model to obtain multi-layer camouflage target representations set sequentially from low-level to high-level;
目标识别模块，由各层伪装目标表征中提取伪装表征分片，并输入至预先训练好的目标识别模型获得相应等级的多个伪装识别特征；The target recognition module extracts camouflage representation fragments from each layer of camouflage target representations, and inputs them into the pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels;
伪装识别图像生成模块，将各伪装识别特征进行卷积获得伪装识别层级图像，将各伪装识别层级图像进行叠加获得伪装识别图像；The camouflage recognition image generation module convolves each camouflage recognition feature to obtain a camouflage recognition level image, and superimposes each camouflage recognition level image to obtain a camouflage recognition image;
训练模块，用于训练所述目标搜索模型和目标识别模型。A training module for training the target search model and target recognition model.
优选的，目标搜索模型包括依次设置特征提取层Stage1、特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5；所述特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5依次对采集图像提取伪装表征f；所述特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5分别连接有用于对伪装表征f的局部细节信息放大的感受野网络；Preferably, the target search model includes sequentially setting the feature extraction layer Stage1, the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5; the feature extraction layer Stage2, the feature extraction layer Stage3, and the feature extraction layer Stage4 and the feature extraction layer Stage5 sequentially extract the camouflage representation f from the collected images; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 are respectively connected to amplify the local detail information of the camouflage representation f. wild network;
所述目标识别模型包括用于小波注意力变换的四个小波注意力单元；所述四个小波注意力单元与四个感受野网络一一对应连接。The target recognition model includes four wavelet attention units for wavelet attention transformation; the four wavelet attention units are connected with the four receptive field networks in one-to-one correspondence.
本发明第三方面提供了计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现所述伪装目标检测方法的步骤。A third aspect of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the steps of the camouflaged target detection method.
与现有技术相比，本发明的有益效果：Compared with the prior art, the beneficial effects of the present invention:
(1)本发明采集的检测图像输入至预先训练好的目标搜索模型，获得低级到高级依次设置的多层伪装目标表征；目标搜索模型内设有感受野网络，通过感受野网络对提取的伪装表征f进行感受野放大，有利于提高识别率的准确性。(1) The detection images collected by the present invention are input into the pre-trained target search model to obtain multi-layer camouflage target representations set in sequence from low to high levels; the target search model is provided with a receptive field network, and the extracted camouflage It is beneficial to improve the accuracy of the recognition rate by characterizing f to enlarge the receptive field.
(2)本发明由各层伪装目标表征中提取伪装表征分片，并输入至预先训练好的目标识别模型获得相应等级的多个伪装识别特征；通过多尺度提取伪装表征分片输入目标识别模型中，重新拼接出伪装表征，捕获全局局部的上下文细节信息，可以更好识别边缘纹理结构，从而获得更高的识别准确率。(2) The present invention extracts camouflage representation fragments from each layer of camouflage target representation, and inputs them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels; extracts camouflage representation fragments through multiple scales and inputs the target recognition model In , the camouflage representation is re-spliced and the global and local contextual details are captured, which can better identify the edge texture structure and obtain a higher recognition accuracy.
(3)本发明将各伪装识别特征进行卷积获得伪装识别层级图像，将各伪装识别层级图像进行叠加获得伪装识别图像；通过基于小波学习网络的目标识别模型来有效检测出伪装目标，提高识别率的准确性。(3) In the present invention, each camouflage recognition feature is convolved to obtain a camouflage recognition level image, and each camouflage recognition level image is superimposed to obtain a camouflage recognition image; the camouflage target is effectively detected through the target recognition model based on the wavelet learning network, and the recognition is improved. rate accuracy.
附图说明Description of drawings
图1是本发明实施例提供的一种基于步态信息的疲劳检测方法的流程图；1 is a flowchart of a method for detecting fatigue based on gait information provided by an embodiment of the present invention;
图2是本发明实施例提供的感受野网络的结构图；2 is a structural diagram of a receptive field network provided by an embodiment of the present invention;
图3是本发明实施例提供的各方法识别伪装目标的视觉对比图；3 is a visual comparison diagram of identifying a camouflaged target by each method provided by an embodiment of the present invention;
图4是本发明实施例提供的小波注意力单元的结构图。FIG. 4 is a structural diagram of a wavelet attention unit provided by an embodiment of the present invention.
具体实施方式Detailed ways
下面结合附图对本发明作进一步描述。以下实施例仅用于更加清楚地说明本发明的技术方案，而不能以此来限制本发明的保护范围。The present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solutions of the present invention more clearly, and cannot be used to limit the protection scope of the present invention.
实施例一Example 1
如图1所示，本发明第一方面提供了一种基于学习神经网络的伪装目标检测方法，包括：As shown in FIG. 1 , a first aspect of the present invention provides a method for detecting a camouflaged target based on a learning neural network, including:
所述采集检测图像输入至预先训练好的目标搜索模型，获得低级到高级依次设置的多层伪装目标表征的方法包括：The method of collecting and detecting images and inputting them into a pre-trained target search model, and obtaining multi-layer camouflage target representations set in sequence from low-level to high-level includes:
将检测图像输入至目标搜索模型中的Res2Net50残差网络，所述Res2Net50残差网络包括特征提取层Stage1、特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5，涵盖了从高分辨率、弱语义到低分辨率、强语义的多样化特征金字塔，所述特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5输出低级到高级依次设置的四层伪装表征f；Input the detection image to the Res2Net50 residual network in the target search model, the Res2Net50 residual network includes the feature extraction layer Stage1, the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5, covering from High-resolution, weak semantic to low-resolution, strong semantic diverse feature pyramid, the feature extraction layer Stage2, feature extraction layer Stage3, feature extraction layer Stage4 and feature extraction layer Stage5 output four layers of camouflage set in sequence from low level to high level characterize f;
通过感觉野网络对各层伪装表征f的局部细节信息放大，分别获得伪装目标表征fr2、伪装目标表征fr3、伪装目标表征fr4和伪装目标表征fr5。所述感受野模块是一个多分支卷积块，它的内部结构可以分为如下部分：具有不同内核的多分支卷积层和尾随扩张卷积层。分支卷积层负责模拟多尺度的感受野功能，尾随扩张卷积层再现人类视觉系统中感受野尺寸与偏心率的关系。Through the sensory field network, the local detail information of each layer of camouflage representation f is amplified, and the camouflage target representation fr2 , the camouflage target representation fr3 , the camouflage target representation fr4 and the camouflage target representation fr5 are obtained respectively. The receptive field module is a multi-branch convolution block, and its internal structure can be divided into the following parts: multi-branch convolution layers with different kernels and trailing dilated convolution layers. The branched convolutional layer is responsible for simulating the multi-scale receptive field function, and the trailing dilated convolutional layer reproduces the relationship between the receptive field size and eccentricity in the human visual system.
其中，如图2所示，感受野模块包括五个分支。第一分支为卷积1×1；第二分支依次设置有卷积1×1、卷积1×3、卷积3×1和空洞卷积×3；第三分支依次设置有卷积1×1、卷积1×5、卷积5×1和空洞卷积×5；第二分支依次设置有卷积1×1、卷积1×7、卷积7×1和空洞卷积×7；前四个分支串联后，通过1×1卷积层操作，其通道降为既定好的通道数。最后，加入第五个分支，并整体输入进激活函数已获得增强后的特征。Among them, as shown in Figure 2, the receptive field module includes five branches. The first branch is
由各层伪装目标表征中提取伪装表征分片的方法包括：The methods of extracting camouflage representation fragments from camouflage target representations of each layer include:
提取伪装目标表征fr5的四分片作为伪装表征分片fr′5；提取伪装目标表征fr4的二分片作为伪装表征分片fr′4；提取伪装目标表征fr3的整体作为伪装表征分片fr′3；提取伪装目标表征fr2的整体作为伪装表征分片fr′2。extracting the quadrant of the camouflage target representation fr5 as the camouflage representation segment fr′5 ; extracting the second segment of the camouflage target representation fr4 as the camouflage representation segment fr′4 ; extracting the whole of the camouflage target representation fr3 as the camouflage representation segment fr′3 ; extract the whole of the camouflage target representation fr2 as the camouflage representation fragment fr′2 .
伪装表征分片输入至预先训练好的目标识别模型获得相应等级的多个伪装识别特征的方法包括：The methods for obtaining multiple camouflage recognition features of corresponding levels by inputting the camouflage representation into slices into the pre-trained target recognition model include:
按照由低级到高级的顺序将伪装表征分片fr′5、伪装表征分片fr′4、伪装表征分片fr′3和伪装表征分片fr′2依次输入至目标识别模型；Input the camouflage representation segment fr'5 , the camouflage representation segment fr'4 , the camouflage representation segment fr'3 and the camouflage representation segment fr'2 into the target recognition model in order from low level to high level;
对最低级的伪装表征分片fr′5进行小波注意力变换提取伪装识别特征r5；Perform wavelet attention transform on the lowest level camouflage representation fragment fr'5 to extract camouflage identification feature r5 ;
将下一级伪装表征分片rfn和上一级提取的伪装识别特征rn+1结合后，通过小波注意力变换依次获得伪装识别特征，具体方法包括：After combining the next-level camouflage representation fragment rfn and the camouflage recognition feature rn+1 extracted by the previous level, the camouflage recognition features are sequentially obtained through wavelet attention transformation, and the specific methods include:
本发明采用小波变换来分解频域中的特征。通过加法运算和转置卷积重建频率子带，对于2×2块中的像素(每一块分别表示为a、b、c和d)，二维哈尔小波系数的计算过程定义为：The present invention uses wavelet transform to decompose the features in the frequency domain. The frequency subbands are reconstructed by addition operation and transposed convolution. For pixels in 2 × 2 blocks (each block is denoted as a, b, c, and d, respectively), the calculation process of the two-dimensional Haar wavelet coefficients is defined as:
其中，A代表低频信息。B、C、D分别代表水平、垂直和对角线方向的高频，分解后的高度和宽度子带是原始图像大小的一半。然后分别增强高频子带和低频子带。其中高频信息主要包含纹理和边缘信息，提取高频信息并连接解码器已保留类似边缘的信息。此外，使用一个带有注意力模块的残差块来增强高频自带。具体来说，注意力模块有通道注意力和空间注意力组成。由于高频信息稀疏，用最大池化代替平均池化以此强调重要的通道。然后该模块使得小波学习网络更加关注空间信息征。对于低频信息，我们提出了一种归一化和注意力机制的新型低频增强分支。通过自适应实例归一化将图像转换为任意样式，该层将伪装分割信息(内容表征)的均值和方差与轮廓信息(样式表征)的均值和方差对齐。Among them, A represents the low frequency information. B, C, and D represent high frequencies in the horizontal, vertical, and diagonal directions, respectively, and the decomposed height and width subbands are half the size of the original image. The high frequency sub-band and the low frequency sub-band are then separately enhanced. The high-frequency information mainly includes texture and edge information, and the high-frequency information is extracted and connected to the decoder to preserve the edge-like information. Furthermore, a residual block with an attention module is used to enhance the high-frequency self-contained. Specifically, the attention module consists of channel attention and spatial attention. Since high-frequency information is sparse, max pooling is used instead of average pooling to emphasize important channels. Then this module makes the wavelet learning network pay more attention to the spatial information features. For low-frequency information, we propose a novel low-frequency enhancement branch with normalization and attention mechanism. Images are transformed into arbitrary styles via adaptive instance normalization, a layer that aligns the mean and variance of camouflage segmentation information (content representation) with the mean and variance of contour information (style representation).
如图4所示，将下一级伪装表征分片rfn(n≥1，n表示为伪装目标表征对应的序号)和上一级提取的伪装识别特征rn+1进行结合后进行卷积，conv表示为卷积，获得低频分量LL、高频分量LH、高频分量HL和高频分量HH；As shown in Figure 4, the next-level camouflage representation fragment rfn (n≥1, n is the sequence number corresponding to the camouflage target representation) and the camouflage identification feature rn+1 extracted from the previous level are combined and then convolved , conv is expressed as convolution to obtain low-frequency component LL, high-frequency component LH, high-frequency component HL and high-frequency component HH;
由所述低频分量LL捕获伪装目标的伪装分割信息和；通过注意力机制AM和含有实例归一化残差块ResIN对低频分量LL进行加强，叠加伪装分割信息和轮廓信息，获得低频分量伪装特征rLL；The camouflage segmentation information sum of the camouflage target is captured by the low-frequency component LL; the low-frequency component LL is enhanced through the attention mechanism AM and the residual block ResIN containing the instance normalization, and the camouflage segmentation information and contour information are superimposed to obtain the low-frequency component camouflage feature. rLL ;
将高频分量LH、高频分量HL和高频分量HH分别通过归一化残差块ResB和注意力机制AM进行增强，获得高频分量伪装特征rLH、高频分量伪装特征rHL和高频分量伪装特征rHH；The high-frequency component LH, high-frequency component HL and high-frequency component HH are respectively enhanced by the normalized residual block ResB and the attention mechanism AM to obtain the high-frequency component camouflage feature rLH , the high-frequency component camouflage feature rHL and the high-frequency component camouflage feature r HL . frequency component camouflage feature rHH ;
将低频分量伪装特征rLL、高频分量伪装特征rLH、高频分量伪装特征rHL和高频分量伪装特征rHH进行转置卷积TC，获得伪装识别特征rn。Perform transposed convolution TC on the low-frequency component camouflage feature rLL , the high-frequency component camouflage feature rLH , the high-frequency component camouflage feature rHL and the high-frequency component camouflage featurerHH to obtain the camouflage identification feature rn .
将各伪装识别特征进行卷积获得伪装识别层级图像，将各伪装识别层级图像进行叠加获得伪装识别图像。Convolving each camouflage recognition feature to obtain a camouflage recognition level image, and superimposing each camouflage recognition level image to obtain a camouflage recognition image.
所述目标搜索模型和目标识别模型的训练过程包括：The training process of the target search model and target recognition model includes:
采集包含伪装目标的训练图像，获取训练数据集；Collect training images containing camouflaged targets to obtain training data sets;
将通过训练数据集对目标搜索模型和目标识别模型进行训练，计算训练过程中伪装识别的结构化损失值S，根据结构化损失值S迭代调整目标搜索模型和目标识别模型的参数，获得设定要求的目标搜索模型和目标识别模型。The target search model and target recognition model will be trained through the training data set, the structural loss value S of camouflage recognition during the training process will be calculated, and the parameters of the target search model and target recognition model will be iteratively adjusted according to the structural loss value S, and the set value will be obtained. The required target search model and target recognition model.
计算训练过程中伪装识别的结构化损失值S，计算公式为：Calculate the structural loss value S of camouflage recognition during the training process, and the calculation formula is:
S＝α*So+(1-α)*SrS=α*So +(1-α)*Sr
So＝μ*OFG+(1-μ)*OBGSo = μ*OFG +(1-μ)*OBG
其中，So表示为面向物体的结构相似性度量；Sr表示为面向区域的结构相似性度量；α表示为权重因子；Wk表示为第k个伪装表征分片的权值；K表示为伪装表征分片的总量；k表示伪装表征分片的序列号；μ为前景面积所占图像的比例；OBG表示为背景度量；OFG表示为前景度量。Among them, So is the object-oriented structural similarity measure; Sr is the region-oriented structural similarity measure; α is the weight factor; Wk is the weight of the k-th camouflage representation slice; K is expressed as The total number of camouflage representation fragments; k represents the serial number of camouflage representation fragments; μ is the proportion of the foreground area in the image; OBG represents the background metric; OFG represents the foreground metric.
计算所述背景度量OBG和前景度量OFG，表达公式为：Calculate the background metric OBG and foreground metric OFG , and the expression formula is:
其中，DFG表示为伪装预测图与伪装表征分片的前景区域差别程度，，xFG表示伪装目标预测图的前景区域的概率值，yFG分别表示伪装表征分片的前景区域的概率值，
如附图3所示，对于五种主流的目标分割网络模型SINet(Search IdentificationNetwork)、Parallel Reverse Attention Network(PraNet)、CPD(Cascaded PartialDecoder)、PFANet(Pyramidal Feature Aggregation Network)、EGNet(Edge GuidanceNetwork)的直观视觉对比，其中GT(GroundTruth)表示分割真值，Ours表示本发明的伪装目标检测方法；通过对比可知本发明可以恢复出更加清晰的伪装目标图像，同时具备更好的边缘纹理信息，能更好得从与伪装目标呈高相似性的环境中检测出来。As shown in Figure 3, for five mainstream target segmentation network models SINet (Search Identification Network), Parallel Reverse Attention Network (PraNet), CPD (Cascaded PartialDecoder), PFANet (Pyramidal Feature Aggregation Network), EGNet (Edge Guidance Network) Intuitive visual comparison, in which GT (GroundTruth) represents the true value of segmentation, and Ours represents the camouflage target detection method of the present invention; through the comparison, it can be seen that the present invention can restore a clearer camouflage target image, and has better edge texture information at the same time. Good enough to be detected from environments with high similarity to camouflaged targets.
实施例二
本发明第二方面提供了一种基于学习神经网络的伪装目标检测系统，本实施例提供的伪装目标检测系统可以应用于实施例一所述伪装目标检测方法，包括：A second aspect of the present invention provides a camouflaged target detection system based on a learning neural network. The camouflaged target detection system provided in this embodiment can be applied to the camouflaged target detection method described in
采集模块，采集检测图像；The acquisition module collects inspection images;
目标搜索模块，用于将采集的检测图像输入至预先训练好的目标搜索模型，获得低级到高级依次设置的多层伪装目标表征；The target search module is used to input the collected detection images into the pre-trained target search model to obtain multi-layer camouflage target representations set sequentially from low-level to high-level;
目标识别模块，由各层伪装目标表征中提取伪装表征分片，并输入至预先训练好的目标识别模型获得相应等级的多个伪装识别特征；The target recognition module extracts camouflage representation fragments from each layer of camouflage target representations, and inputs them into the pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels;
伪装识别图像生成模块，将各伪装识别特征进行卷积获得伪装识别层级图像，将各伪装识别层级图像进行叠加获得伪装识别图像；The camouflage recognition image generation module convolves each camouflage recognition feature to obtain a camouflage recognition level image, and superimposes each camouflage recognition level image to obtain a camouflage recognition image;
训练模块，用于训练所述目标搜索模型和目标识别模型。A training module for training the target search model and target recognition model.
所述目标搜索模型包括依次设置特征提取层Stage1、特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5；所述特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5依次对采集图像提取伪装表征f；所述特征提取层Stage2、特征提取层Stage3、特征提取层Stage4和特征提取层Stage5分别连接有用于对伪装表征f的局部细节信息放大的感受野网络；所述目标识别模型包括用于小波注意力变换的四个小波注意力单元；所述四个小波注意力单元与四个感受野网络一一对应连接；The target search model includes sequentially setting the feature extraction layer Stage1, the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and The feature extraction layer Stage5 sequentially extracts the camouflage representation f from the collected images; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4, and the feature extraction layer Stage5 are respectively connected with receptive fields used for amplifying the local detail information of the camouflage representation f. network; the target recognition model includes four wavelet attention units for wavelet attention transformation; the four wavelet attention units are connected with the four receptive field networks in one-to-one correspondence;
按照由低级到高级的顺序将伪装表征分片fr′5、伪装表征分片fr′4、伪装表征分片fr′3和伪装表征分片fr′2依次输入至四个小波注意力单元；Input the camouflage representation segment fr'5 , the camouflage representation segment fr'4 , the camouflage representation segment fr'3 and the camouflage representation segment fr'2 into the four wavelet attention units in order from low level to high level;
所述小波注意力单元对最低级的伪装表征分片fr′5进行小波注意力变换提取伪装识别特征r5；The wavelet attention unit performs wavelet attention transformation on the lowest-level camouflage representation fragment fr'5 to extract the camouflage identification feature r5 ;
将下一级伪装表征分片rfn和上一级提取的伪装识别特征rn+1结合后，通过所述小波注意力单元进行小波注意力变换依次获得伪装识别特征。After combining the next-level camouflage representation segment rfn and the camouflage identification feature rn+1 extracted by the upper-level, the camouflage identification features are sequentially obtained by performing wavelet attention transformation through the wavelet attention unit.
实施例三
本发明第三方面提供了计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现实施例一所述伪装目标检测方法的步骤。A third aspect of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the steps of the disguised target detection method described in
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.
以上所述仅是本发明的优选实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明技术原理的前提下，还可以做出若干改进和变形，这些改进和变形也应视为本发明的保护范围。The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the technical principle of the present invention, several improvements and modifications can also be made. These improvements and modifications It should also be regarded as the protection scope of the present invention.
Claims (10)
Translated fromChinese
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210257144.XACN114842324B (en) | 2022-03-16 | 2022-03-16 | Camouflage target detection method and system based on learning neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210257144.XACN114842324B (en) | 2022-03-16 | 2022-03-16 | Camouflage target detection method and system based on learning neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114842324Atrue CN114842324A (en) | 2022-08-02 |
| CN114842324B CN114842324B (en) | 2025-01-17 |
Family
ID=82561999
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210257144.XAActiveCN114842324B (en) | 2022-03-16 | 2022-03-16 | Camouflage target detection method and system based on learning neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114842324B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115830503A (en)* | 2022-12-05 | 2023-03-21 | 中建八局西北建设有限公司 | Method and system for judging illegal water adding behavior of concrete tank truck |
| CN116563684A (en)* | 2023-05-06 | 2023-08-08 | 中国人民解放军军事科学院系统工程研究院 | A light-weight camouflaged target detection method and system |
| CN119418256A (en)* | 2024-10-21 | 2025-02-11 | 中国人民解放军国防科技大学 | A method for detecting camouflaged objects |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751134A (en)* | 2019-12-23 | 2020-02-04 | 长沙智能驾驶研究院有限公司 | Target detection method, storage medium and computer device |
| JP6830707B1 (en)* | 2020-01-23 | 2021-02-17 | 同▲済▼大学 | Person re-identification method that combines random batch mask and multi-scale expression learning |
| CN113449727A (en)* | 2021-07-19 | 2021-09-28 | 中国电子科技集团公司第二十八研究所 | Camouflage target detection and identification method based on deep neural network |
| CN113468996A (en)* | 2021-06-22 | 2021-10-01 | 广州大学 | Camouflage object detection method based on edge refinement |
| CN113793272A (en)* | 2021-08-11 | 2021-12-14 | 东软医疗系统股份有限公司 | Image noise reduction method and device, storage medium and terminal |
- 2022
- 2022-03-16CNCN202210257144.XApatent/CN114842324B/enactiveActive
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751134A (en)* | 2019-12-23 | 2020-02-04 | 长沙智能驾驶研究院有限公司 | Target detection method, storage medium and computer device |
| JP6830707B1 (en)* | 2020-01-23 | 2021-02-17 | 同▲済▼大学 | Person re-identification method that combines random batch mask and multi-scale expression learning |
| CN113468996A (en)* | 2021-06-22 | 2021-10-01 | 广州大学 | Camouflage object detection method based on edge refinement |
| CN113449727A (en)* | 2021-07-19 | 2021-09-28 | 中国电子科技集团公司第二十八研究所 | Camouflage target detection and identification method based on deep neural network |
| CN113793272A (en)* | 2021-08-11 | 2021-12-14 | 东软医疗系统股份有限公司 | Image noise reduction method and device, storage medium and terminal |
Non-Patent Citations (1)
| Title |
|---|
| 贾其等: "人眼视觉注意机制用于伪装效果评价", 应用科学学报, no. 03, 30 May 2011 (2011-05-30)* |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115830503A (en)* | 2022-12-05 | 2023-03-21 | 中建八局西北建设有限公司 | Method and system for judging illegal water adding behavior of concrete tank truck |
| CN116563684A (en)* | 2023-05-06 | 2023-08-08 | 中国人民解放军军事科学院系统工程研究院 | A light-weight camouflaged target detection method and system |
| CN119418256A (en)* | 2024-10-21 | 2025-02-11 | 中国人民解放军国防科技大学 | A method for detecting camouflaged objects |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114842324B (en) | 2025-01-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yeh et al. | Multi-scale deep residual learning-based single image haze removal via image decomposition | |
| CN114842324A (en) | Method and system for detecting disguised target based on learning neural network | |
| CN107657217B (en) | Infrared and visible light video fusion method based on moving target detection | |
| Long et al. | Underwater forward-looking sonar images target detection via speckle reduction and scene prior | |
| CN112750140A (en) | Disguised target image segmentation method based on information mining | |
| Singh et al. | Wavelet based multi class image classification using neural network | |
| CN109102475B (en) | Image rain removing method and device | |
| CN107909109A (en) | SAR image sorting technique based on conspicuousness and multiple dimensioned depth network model | |
| CN111179208B (en) | Infrared-visible light image fusion method based on saliency map and convolutional neural network | |
| Verma et al. | Wild animal detection from highly cluttered images using deep convolutional neural network | |
| CN110084181B (en) | A method for ship target detection in remote sensing images based on sparse MobileNetV2 network | |
| CN114266894B (en) | Image segmentation method, device, electronic device and storage medium | |
| CN116188918A (en) | Image denoising method, training method of network model, device, medium and equipment | |
| Sun et al. | A priori-guided multi-layer rain-aware network for single image deraining | |
| CN112115871A (en) | High and low frequency interweaving edge feature enhancement method suitable for pedestrian target detection and method for building enhanced network | |
| CN119942137A (en) | Lightweight target detection method, lightweight target detection model method and device | |
| CN104361571A (en) | Infrared and low-light image fusion method based on marginal information and support degree transformation | |
| CN106650613A (en) | Deep learning and trunk extraction-based pedestrian identification system and method | |
| CN102521603B (en) | Method for classifying hyperspectral images based on conditional random field | |
| CN110298808A (en) | The domain NSCT infrared image self-adaptive solution method based on quantum differential evolution algorithm | |
| CN106650678B (en) | Gabor Wavelet Subband Correlation Structure Face Recognition Method | |
| Jyothy et al. | Texture-based multiresolution steganalytic features for spatial image steganography | |
| Contreras et al. | Construction of a data integration platform for the passive monitoring of the Antillean manatee in Panama | |
| Nandini et al. | Wavelet-based method for enhancing the visibility of hazy images using Color Attenuation Prior | |
| Ye et al. | Quantitative effects of discrete wavelet transforms and wavelet packets on aerial digital image denoising |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |