WO2019148309A1 - 基于结构信息的红外小目标图像的快速重构方法及系统 - Google Patents
基于结构信息的红外小目标图像的快速重构方法及系统 Download PDFInfo
- Publication number
- WO2019148309A1 WO2019148309A1 PCT/CN2018/074506 CN2018074506W WO2019148309A1 WO 2019148309 A1 WO2019148309 A1 WO 2019148309A1 CN 2018074506 W CN2018074506 W CN 2018074506W WO 2019148309 A1 WO2019148309 A1 WO 2019148309A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- matrix
- target image
- row
- small target
- infrared small
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
Definitions
- the invention belongs to the field of image processing, and in particular relates to a method and a system for quickly reconstructing an infrared small target image based on structural information.
- Infrared small target images are very important and widely used in military and civilian applications. Many targets, such as drones and various general-purpose aircrafts, appear as small targets or weak targets in the images formed by infrared detectors.
- the application of Internet-based infrared images is the core technology in many fields. With the increasing application of infrared technology in various fields, how to realize the reliable transmission of infrared images under the limited bandwidth of the Internet system has become a bottleneck problem in the current network application of infrared technology.
- the traditional data compression technology is based on the Nyquist sampling theorem.
- the complexity is higher at the encoding end and the complexity at the decoding end is lower.
- this is usually contrary to the hardware configuration of the system.
- the sensor node itself is a cheap and energy-saving device, but it has to complete complex computing tasks such as data collection and compression; at the decoding end, it is usually a large and efficient processing device, but only needs to be relatively simple. Calculation task.
- this contradiction is more acute and becomes a key problem that is difficult to overcome in wireless sensor networks.
- the ever-increasing data rate also puts higher demands on Nyquist's law, making hardware design more difficult.
- CS Compressive Sensing
- the technical problem to be solved by the present invention is to provide a fast reconstruction method and system for infrared small target images based on structural information, aiming at solving the problem of simply using compressed sensing and block sparse Bayesian learning methods in the prior art.
- the problem is that the image obtained by the reconstruction has low precision and the reconstruction time is slow.
- the present invention is implemented in this way, a method for quickly reconstructing an infrared small target image based on structural information, comprising:
- Step A acquiring a target in the infrared small target image, and determining a size prior information according to the target;
- Step B sampling the infrared small target image according to the size prior information, and obtaining several sub-images
- Step C performing row stacking on each of the sub-images to obtain a stacking matrix, where the stacking matrix satisfies a multi-measurement vector model
- Step D performing compression measurement on each of the sub-images to obtain a compressed observation matrix including the stacking matrix
- step E the compressed observation matrix is reconstructed by using a block sparse Bayesian learning method to obtain a reconstructed infrared small target image.
- step C specifically includes:
- the line vector formed by the element of the ith row of the infrared small target image X is represented by X i ⁇ , and the step D specifically includes:
- Step D1 for each of the sub-images Compressed measurements are performed in rows, respectively, and line compression observations are obtained by row compression measurements.
- Step D3 performing compression measurement on each column vector y l to obtain a column compressed observation value
- the perceptual matrix obtained by the compression measurement has the same perceptual matrix ⁇ in each column perceptual process;
- Step D4 constructing an observation compression matrix Y, Then there are:
- Step D5 stacking matrix Substituting into equation (5), the final compressed observation matrix Y is obtained, namely:
- step E specifically includes:
- Step E1 will Substituting into equation (6), the simplified compressed observation matrix Y is obtained, then:
- step E2 considering the unknown noise vector V, the formula (7) is converted to obtain:
- step E3 the formula (8) is converted into a block sparse model form, and:
- T represents the transpose of the matrix
- I L represents the unit matrix of the L rows
- ⁇ i represents a hyperparameter, which is used to determine whether the value of the i-th block is zero
- B i represents a positive definite matrix, which is used to model the structural features between the elements in the i-th block
- step E5 the prior of the block sparse signal x is:
- ⁇ diag ( ⁇ 1 , ⁇ 2 , ..., ⁇ SN );
- ⁇ i is the i-th column of the matrix ⁇ .
- the embodiment of the invention further provides a fast reconstruction system for infrared small target images based on structural information, comprising:
- a prior acquisition unit configured to acquire a target in the infrared small target image, and determine a size prior information according to the target;
- An image sampling unit configured to sample the infrared small target image according to the size prior information, to obtain a plurality of sub-images
- a row stacking unit configured to perform row stacking on each of the sub-images to obtain a stacking matrix, wherein the stacking matrix satisfies a multi-measurement vector model
- a compression measuring unit configured to perform compression measurement on each of the sub-images to obtain a compressed observation matrix including the stacked matrix
- An image reconstruction unit is configured to reconstruct the compressed observation matrix by using a block sparse Bayesian learning method to obtain a reconstructed infrared small target image.
- row stacking unit is specifically configured to:
- the row vector formed by the elements of the infrared small target image X and the i-th row is represented by X i ⁇ , and the compression measurement unit is specifically configured to perform the following steps:
- Step D1 for each of the sub-images Compressed measurements are performed in rows, respectively, and line compression observations are obtained by row compression measurements.
- Step D3 performing compression measurement on each column vector y l to obtain a column compressed observation value
- the perceptual matrix obtained by the compression measurement has the same perceptual matrix ⁇ in each column perceptual process;
- Step D4 constructing an observation compression matrix Y, Then there are:
- Step D5 stacking matrix Substituting into equation (20), the final compressed observation matrix Y is obtained, namely:
- image reconstruction unit is specifically configured to perform the following steps:
- Step E1 will Substituting into equation (21), the simplified compressed observation matrix Y is obtained, then:
- step E2 considering the unknown noise vector V, the formula (22) is converted to obtain:
- step E3 the formula (23) is converted into a block sparse model form, and:
- T represents the transpose of the matrix
- I L represents the unit matrix of the L rows
- ⁇ i represents a hyperparameter, which is used to determine whether the value of the i-th block is zero
- B i represents a positive definite matrix, which is used to model the structural features between the elements in the i-th block
- step E5 the prior of the block sparse signal x is:
- ⁇ diag ( ⁇ 1 , ⁇ 2 , ..., ⁇ SN );
- ⁇ i is the i-th column of the matrix ⁇ .
- the present invention has the beneficial effects that the embodiment of the present invention obtains the target in the infrared small target image, determines the size prior information according to the target, and samples the infrared small target image according to the size prior information.
- a tow image is obtained, and each of the dry sub-images is stacked in a row to obtain a stacked matrix.
- the stacked matrix satisfies a multi-measurement vector model, and each sub-image is subjected to compression measurement to obtain a compressed sensing value including a stacked matrix, and finally a block sparse shell is used.
- the Yesi learning method reconstructs the compressed observation matrix to obtain a reconstructed infrared small target image.
- the embodiment of the invention combines the block sparse Bayesian compression sensing method for image reconstruction, and can accurately reconstruct the single frame target image, so that the restored image has a higher average peak signal to noise ratio.
- the embodiment uses a multi-measurement vector model to model the correlation between rows, which reduces the calculation amount of the reconstruction process and improves the running speed of the algorithm.
- FIG. 1 is a flowchart of a method for quickly reconstructing an infrared small target image based on structure information according to an embodiment of the present invention
- FIG. 2 is a schematic diagram of a modeling process of correlation between rows of infrared small target images according to an embodiment of the present invention
- FIG. 5 is a morphologically filtered single target infrared image according to an embodiment of the present invention.
- FIG. 7 is a schematic structural diagram of a fast reconstruction system for an infrared small target image based on structure information according to an embodiment of the present invention.
- FIG. 1 is a schematic diagram of a method for quickly reconstructing an infrared small target image based on structure information according to an embodiment of the present invention, including:
- S101 Acquire an object in an infrared small target image, and determine a size prior information according to the target;
- the infrared small target image is sampled according to the size prior information, and several sub-images are obtained.
- S105 Reconstruct the compressed observation matrix by using a block sparse Bayesian learning method to obtain a reconstructed infrared small target image.
- the embodiment of the invention is based on the compressed sensing technology, and adopts the Bayesian method to introduce the structural prior information of the infrared small target image, thereby promoting the sparse decomposition of the infrared image data and improving the performance of the compression and reconstruction algorithm.
- the compressed sensing method combining the inter-line structure information of the small target image is referred to as a Reshaped Bayesian Compressive Sensing (R-BCS) method, and the following sections will be This method is elaborated.
- the observation g is a signal with a lower degree of freedom than the signal f.
- Restoring the signal f from the observed g is a solution to the ill-conditioned equation, and there are infinite sets of solutions.
- the decomposition coefficient w has a sparse property, so that the number of unknowns is greatly reduced, and the solution of the equation becomes possible.
- the infrared small target image has a certain sparsity. According to the sparse characteristics of the infrared small target image, the embodiment of the invention re-forms the array of rows and columns, and then performs compressed sensing on the reconstructed new array, thereby achieving the purpose of reducing compression time and improving compression efficiency.
- the infrared small target image is X ij represents the element of the i-th row and the j-th column of the infrared small target image.
- an array of infrared small target images is sampled in the lower L line.
- the row vector formed by the elements of the infrared small target image X i-th row is represented in the form of the symbol X i ⁇ throughout the text.
- each sub-image is first stacked in rows to obtain an L-column vector, and then X d is a newly obtained matrix in which the row stack of each column sub-image is merged, that is, then:
- adjacent columns correspond to adjacent rows of the infrared small target image X
- Y is a compression observation finally obtained after the image is reformed.
- the obtained matrix X d is a plurality of measurement vectors having a common sparse hypothesis.
- the matrix is formed in a form that satisfies the Multiple Measurement Vectors (MMV).
- MMV Multiple Measurement Vectors
- the traditional multi-measure vector compression sensing method assumes that the sparse structure of the sparse signal does not change with time.
- the reconstruction method provided by the embodiment of the present invention models the inter-row structure of the signal of the infrared small target image according to the idea of modeling the time structure of the same support set timing signal according to the MMV model.
- the multi-frame sparse signal recovery problem with temporal correlation is changed to the single-frame sparse image signal recovery problem using the inter-row structure prior information.
- V represents an unknown noise vector
- T represents the transpose of the matrix
- I L represents the unit matrix of the L rows
- the MMV model of equation (10) is transformed into a single measurement vector model SMV (Single Measurement Vectors) with a block structure.
- SMV Single Measurement Vectors
- the purpose of this is to facilitate the conversion of the inter-row structure prior information of the target image into block structure information.
- the block structure information is used in the reconstruction algorithm.
- Block sparse signal x can be written as among them Indicates that it is the i-th block of x.
- the probability density of each block x i is set to obey the Gaussian distribution, namely:
- ⁇ i is a hyperparameter and its value is non-negative, which is used to determine whether the value of the i-th block is zero
- B i is a positive definite matrix, and its value is temporarily unknown, and is used for each element in the i-th block.
- the structural features are modeled.
- the a priori of the block sparse signal x is:
- the positive definite matrix B is used as the final estimate of each B i in order to prevent over-fitting.
- ⁇ i is the i-th column of the matrix ⁇ .
- the embodiment of the invention provides a fast reconstruction method of the infrared image combined with the structural information. Firstly, according to the prior information of the target in the infrared small target image, the infrared small target image is reordered, so that the reconstructed image array form satisfies the MMV model. Secondly, the MMV model is used to construct the signal with the same support set.
- the frame sparse image signal recovers the infrared small target image.
- multiple types of infrared image pairing algorithms are validated and compared with the block sparse Bayesian learning algorithm (BSBL). From the experimental results, the proposed method improves the algorithm running time and reconstruction accuracy compared with the BSBL algorithm.
- BSBL block sparse Bayesian learning algorithm
- the single target infrared image and the multi-target infrared image are taken as experimental objects respectively, and the two technical indicators of the peak signal to noise ratio and the algorithm reconstruction time of the reconstruction algorithm are performed. Verification.
- the images used in the experiment are shown in Figures 3 and 4.
- the image decoding end decodes the foreground image and the background image separately after receiving the data, and the foreground and background are synthesized to restore the original image.
- the reconstruction method provided by the embodiment of the present invention firstly uses the morphological detection method (top-hat) to filter the image, and separates the target image from the background image, and the obtained target image is 5 And Figure 6 shows.
- the image pixel is 320 x 256, and the target size is 8 x 8.
- This image has typical spatial sparsity.
- the infrared small target image is sampled in the next 8 lines.
- the reconstruction method provided by the embodiment of the present invention compresses and then recovers, and the peak signal to noise ratio of the restored image is 106.6191, and the relative error with the original image is 0.1009.
- the pixel sizes of the multiple targets are not consistent, roughly 10 x 7, 9 x 6, and 10 x 6.
- the pixel size of each target is close to 8, which assumes that the small target occupies a pixel width of 8, and the image is sampled in the next 8 lines in the experiment.
- the reconstruction method provided by the embodiment of the present invention compresses and then recovers, and the peak signal to noise ratio of the restored image is 102.1559, and the relative error from the original image is 0.1337.
- the BSBL algorithm is a 20-time Monte Carlo experiment. The specific experimental results are shown in the following table:
- a fast reconstruction method combining the inter-row structure information of the small target image, namely the R-BCS method.
- the method can reconstruct the single-frame target image more accurately, and the average peak signal-to-noise ratio of the restored image is relatively high.
- the fast reconstruction method uses the MMV model to model the inter-row correlation, which reduces the computational complexity of the reconstruction process. , improve the speed of the algorithm.
- the experimental results show that the R-BCS method is used to compress the infrared small target image, and then the image is reconstructed based on the Bayesian method.
- the infrared small target image can be restored with higher precision, and the reconstruction method is much faster than BSBL. algorithm.
- FIG. 7 is a diagram showing a fast reconstruction system of an infrared small target image based on structure information according to an embodiment of the present invention, comprising:
- a prior acquisition unit 701 configured to acquire a target in the infrared small target image, and determine a size prior information according to the target;
- the image sampling unit 702 is configured to sample the infrared small target image according to the size prior information to obtain a plurality of sub-images;
- a row stacking unit 703 configured to perform row stacking on each of the sub-images to obtain a stacking matrix, where the stacking matrix satisfies a multi-measurement vector model;
- a compression measurement unit 704 configured to perform compression measurement on each of the sub-images to obtain a compressed observation matrix including the stacking matrix
- the image reconstruction unit 705 is configured to reconstruct the compressed observation matrix by using a block sparse Bayesian learning method to obtain a reconstructed infrared small target image.
- row stacking unit 703 is specifically configured to:
- the row vector formed by the element of the infrared small target image X and the i-th row is represented by X i ⁇ , and the compression measuring unit 704 is specifically configured to perform the following steps:
- Step D1 for each of the sub-images Compressed measurements are performed in rows, respectively, and line compression observations are obtained by row compression measurements.
- Step D3 performing compression measurement on each column vector y l to obtain a column compressed observation value
- ⁇ indicates the right
- the perceptual matrix obtained by the compression measurement has the same perceptual matrix ⁇ in each column perceptual process
- Step D4 constructing an observation compression matrix Y, Then there are:
- Step D5 stacking matrix Substituting into equation (23), the final compressed observation matrix Y is obtained, namely:
- image reconstruction unit 705 is specifically configured to perform the following steps:
- Step E1 will Substituting into equation (24), the simplified compressed observation matrix Y is obtained, then:
- step E2 considering the unknown noise vector V, the formula (25) is converted to obtain:
- step E3 the formula (26) is converted into a block sparse model form to obtain:
- T represents the transpose of the matrix
- I L represents the unit matrix of the L rows
- ⁇ i represents a hyperparameter, which is used to determine whether the value of the i-th block is zero
- B i represents a positive definite matrix, which is used to model the structural features between the elements in the i-th block
- step E5 the prior of the block sparse signal x is:
- ⁇ diag ( ⁇ 1 , ⁇ 2 , ..., ⁇ SN );
- ⁇ i is the i-th column of the matrix ⁇ .
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Studio Devices (AREA)
- Transforming Light Signals Into Electric Signals (AREA)
Abstract
本发明适用于图像处理领域,提供了一种基于结构信息的红外小目标图像的快速重构方法,包括:获取红外小目标图像中的目标,根据目标确定尺寸先验信息;根据尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;对每一子图像进行行堆叠,得到堆叠矩阵,堆叠矩阵满足多测量向量模型;对每一子图像进行压缩测量,得到包括堆叠矩阵的压缩观测矩阵;采用块稀疏贝叶斯学习方法对压缩观测矩阵进行重构,得到重构后的红外小目标图像。本发明实施例根据红外小目标图像的成像特点,结合块稀疏贝叶斯压缩感知方法进行图像重构,能够精确地重构单帧目标图像,使还原后的图像具有较高平均峰值信噪比,同时降低了重构过程的计算量,提高了算法运行速度。
Description
本发明属于图像处理领域,尤其涉及一种基于结构信息的红外小目标图像的快速重构方法及系统。
红外小目标图像在军事上和民用上应用非常重要且广泛。诸多目标,如无人机、各种通用飞行器等,在红外探测器形成的图像中均表现为小目标或弱小目标。基于互联网的红外图像的应用更是许多领域的核心技术。随着红外技术在各领域的应用越来越广泛,如何实现红外图像在互联网系统有限的带宽下的可靠传输,成为当前红外技术在网络应用中存在的瓶颈问题。
传统的数据压缩技术基于奈奎斯特采样定理,在编码端复杂度较高,解码端复杂度较低,对于无线传感器网络或其他系统来说,这通常与系统的硬件配置相悖。以无线传感器网络为例,传感器节点本身是廉价且节能的设备,却要完成数据的采集和压缩这类复杂的计算任务;在解码端通常是大型高效的处理设备,却仅需完成相对简单的计算任务。在野外作业或者军事作业的场合,这一矛盾更加尖锐,成为无线传感器网络中难以克服的关键问题。而数据率的不断提高,也对奈奎斯特定律提出了更高的要求,从而使得硬件上的设计更为困难。
压缩感知(CS,compressive sensing)方法为这一问题提供了解决方案。压缩感知有别于传统信号采样理论,是一种直接用来对信号中有效信息进行采样压缩的理论框架。针对稀疏图像信号,ZL Zhang等学者对信号中的块状结构信息进行研究和挖掘,形成了块稀疏贝叶斯学习(Block Sparse Bayesian Learning,BSBL)方法理论,该理论针对信号的块内相关性而提出,是一种有效的图像压缩感知方法,已得到众多学者的关注。
但是现有技术单纯地使用压缩感知和块稀疏贝叶斯学习方法进行图像的重构,导致出现重构得到的图像精度低,重构时间慢的问题。
发明内容
本发明所要解决的技术问题在于提供一种基于结构信息的红外小目标图像的快速重构方法及系统,旨在解决现有技术单纯地使用压缩感知和块稀疏贝叶斯学习方法进行图像的重构,导致出现重构得到的图像精度低,重构时间慢的问题。
本发明是这样实现的,一种基于结构信息的红外小目标图像的快速重构方法,包括:
步骤A,获取红外小目标图像中的目标,根据所述目标确定尺寸先验信息;
步骤B,根据所述尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;
步骤C,对每一所述子图像进行行堆叠,得到堆叠矩阵,所述堆叠矩阵满足多测量向量模型;
步骤D,对每一所述子图像进行压缩测量,得到包括所述堆叠矩阵的压缩观测矩阵;
步骤E,采用块稀疏贝叶斯学习方法对所述压缩观测矩阵进行重构,得到重构后的红外小目标图像。
进一步地,所述步骤C具体包括:
进一步地,以X
i·表示所述红外小目标图像X第i行元素所构成的行向量,所述步骤D具体包括:
进一步地,所述步骤E具体包括:
Y=Φ·X
d (7)
步骤E2,考虑未知的噪声向量V,将公式(7)进行转换,得到:
Y=Φ·X
d+V (8)
步骤E3,将公式(8)转换成块稀疏模型形式,得到:
y=D·x+v (9)
设定每一个块x
i的概率密度服从高斯分布,即:
p(x
i)=N(0,γ
iB
i),i=1,…,SN; (10)
其中,γ
i表示超参数,用来决定第i个块的值是否均为零,B
i表示正定矩阵,用来对第i个块里各个元素间的结构特征进行建模;
步骤E5,块稀疏信号x的先验为:
p(x)=N(0,∑); (11)
设公式(9)y=D·x+v中的噪声向量v中每个元素服从高斯分布p(v
i)~N(0,λ);
采用块稀疏贝叶斯学习方法对y=D·x+v进行重构,得到重构后的红外小目标图像:
X
d=ГΦ
T(λI+ΦГΦ
T)
-1Y; (12)
其中:Γ=diag(γ
1,γ
2,…,γ
SN);
正定矩阵B作为每一个B
i的最终估计值:
本发明实施例还提供了一种基于结构信息的红外小目标图像的快速重构系统,包括:
先验获取单元,用于获取红外小目标图像中的目标,根据所述目标确定尺寸先验信息;
图像采样单元,用于根据所述尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;
行堆叠单元,用于对每一所述子图像进行行堆叠,得到堆叠矩阵,所述堆叠矩阵满足多测量向量模型;
压缩测量单元,用于对每一所述子图像进行压缩测量,得到包括所述堆叠矩阵的压缩观测矩阵;
图像重构单元,用于采用块稀疏贝叶斯学习方法对所述压缩观测矩阵进行重构,得到重构后的红外小目标图像。
进一步地,所述行堆叠单元具体用于:
进一步地,以X
i·表示所述红外小目标图像X第i行元素所构成的行向量,所述压缩测量单元具体用于执行以下步骤:
进一步地,所述图像重构单元具体用于执行以下步骤:
Y=Φ·X
d (22)
步骤E2,考虑未知的噪声向量V,将公式(22)进行转换,得到:
Y=Φ·X
d+V (23)
步骤E3,将公式(23)转换成块稀疏模型形式,得到:
y=D·x+v (24)
设定每一个块x
i的概率密度服从高斯分布,即:
p(x
i)=N(0,γ
iB
i),i=1,…,SN; (25)
其中,γ
i表示超参数,用来决定第i个块的值是否均为零,B
i表示正定矩阵,用来对第i个块里各个元素间的结构特征进行建模;
步骤E5,块稀疏信号x的先验为:
p(x)=N(0,∑); (26)
设公式(24)y=D·x+v中的噪声向量v中每个元素服从高斯分布p(v
i)~N(0,λ);
采用块稀疏贝叶斯学习方法对y=D·x+v进行重构,得到重构后的红外小目标图像:
X
d=ГΦ
T(λI+ΦГΦ
T)
-1Y; (27)
其中:Γ=diag(γ
1,γ
2,…,γ
SN);
正定矩阵B作为每一个B
i的最终估计值:
本发明与现有技术相比,有益效果在于:本发明实施例通过获取红外小目标图像中的目标,根据该目标确定尺寸先验信息,根据该尺寸先验信息对该红外小目标图像进行采样得到拖杆子图像,对每一干子图像进行行堆叠,得到堆叠矩阵,该堆叠矩阵满足多测量向量模型,对每一子图像进行压缩测量,得 到包含堆叠矩阵的压缩感知值,最后采用块稀疏贝叶斯学习方法对该压缩观测矩阵进行重构,得到重构后的红外小目标图像。本发明实施例根据红外小目标图像的成像特点,结合块稀疏贝叶斯压缩感知方法进行图像重构,能够精确地重构单帧目标图像,使还原后的图像具有较高平均峰值信噪比,同时本实施例采用多测量向量模型对行间相关性进行建模,降低了重构过程的计算量,提高了算法运行速度。
图1是本发明实施例提供的一种基于结构信息的红外小目标图像的快速重构方法的流程图;
图2是本发明实施例提供的红外小目标图像行间相关性的建模过程示意图;
图3是本发明实施例提供的单目标红外图像;
图4是本发明实施例提供的多目标红外图像;
图5是本发明实施例提供的形态学滤波后的单目标红外图像;
图6是本发明实施例提供的形态学滤波后的多目标红外图像;
图7是本发明实施例提供的一种基于结构信息的红外小目标图像的快速重构系统的结构示意图。
为了使本发明的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本发明进行进一步详细说明。应当理解,此处所描述的具体实施例仅仅用以解释本发明,并不用于限定本发明。
图1示出了本发明实施例提供的一种基于结构信息的红外小目标图像的快速重构方法,包括:
S101,获取红外小目标图像中的目标,根据所述目标确定尺寸先验信息;
S102,根据所述尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;
S103,对每一所述子图像进行行堆叠,得到堆叠矩阵,所述堆叠矩阵满足多测量向量模型;
S104,对每一所述子图像进行压缩测量,得到包括所述堆叠矩阵的压缩观测矩阵;
S105,采用块稀疏贝叶斯学习方法对所述压缩观测矩阵进行重构,得到重构后的红外小目标图像。
本发明实施例基于压缩感知技术,采用贝叶斯方法引入红外小目标图像的结构先验信息,以此促进红外图像数据的稀疏分解,提高压缩和重构算法的性能。本发明实施例中将这种结合小目标图像行间结构信息的压缩感知方法称作小目标图像重整贝叶斯压缩感知(Reshaped Bayesian Compressive Sensing,R-BCS)方法,下述各节将对该方法进行详细阐述。
一、压缩感知基本理论
设有N维实信号f,f∈R
N,存在某组正交基B,B∈R
N×N,可将信号f在该正交基下展开,f=Bw,w∈R
N,w表示信号f在正交基B下的分解系数。若w中非零项的系数个数为K,且K<<N,则称w是K-稀疏的。对大多数自然信号,经小波基、DCT基等正交基分解后,其w中大部分分量系数均可忽略。通过设置阈值将这些可忽略的分量置零,即可达到压缩信号f的目的。
在压缩感知技术中,通过一个随机投影矩阵,获得信号f的观测g:g=Φ
f=ΦB
Tw,其中,Φ=[r
1,r
2,...,r
M]
T,为M×N维矩阵,M<N。显然,观测g是一个比信号f自由度低的信号。从观 测g中恢复信号f,是一个病态方程的求解问题,存在无穷多组解。但分解系数w具有稀疏特性,使得未知数的个数大为减少,方程的求解成为可能。业已证明,只要信号在正交基字典B下可进行稀疏分解,且由正交基字典和观测矩阵所确定的算子A=ΦB满足任意2K列(K为信号f经正交基B分解后的稀疏度)均线性无关,即可通过下式中的l
1-范数来求解系数w:
二、基于结构信息的红外小目标图像阵列重整
由红外成像原理可知,红外小目标图像具有一定的稀疏性。本发明实施例依据红外小目标图像的稀疏特性,对其行列阵列进行重整,再对重整后的新阵列进行压缩感知,从而达到降低压缩时间、提高压缩效率的目的。
假设红外小目标图像为
X
ij代表该红外小目标图像第i行第j列的元素。在具体应用中,以目标在该红外小目标图像中的尺寸为先验信息,假设该目标的大小为L×L,设M可被L整除,即M=L×S,重整压缩过程具体包含四大步骤,分别包括:
在全文中以符号X
i·的形式来代表红外小目标图像X第i行元素所构成的行向量。
在矩阵X
d中,相邻列对应于所述红外小目标图像X的相邻行,Y为对图像经过重整后最终所得的压缩观测值。
由步骤可知,对空域稀疏的红外小目标图像经上述步骤完成压缩测量后,所获矩阵X
d是由具有公共稀疏假设的多个测量向量
构成的矩阵,其形式满足多测量向量模型MMV(Multiple Measurement Vectors)。对红外小目标图像的图像阵列的行列进行重排的过程如图2所示。
图像重整贝叶斯压缩感知恢复
传统的多测量向量压缩感知方法是假设稀疏信号的稀疏结构不随时间变化。根据这一特点,本发明实施例提供的重构方法根据MMV模型对具有相同支撑集时序信号的时间结构建模的思路,来对红外小目标图像的信号的行间结构建模。为利用MMV模型解决红外小目标图像的恢复,本实施例中将具有时间相关性的多帧稀疏信号恢复问题变为利用行间结构先验信息的单帧稀疏图像信号恢复问题。
则公式(7)可以简化为:
Y=Φ·X
d (9)
在实际应用,在考虑噪声向量的情况下,公式(9)的模型应为:
Y=Φ·X
d+V (10)
其中V代表未知的噪声向量。
将上述公式(10)所代表的模型转换为块稀疏模型形式,即:
y=D·x+v (11)
公式(10)的MMV模型被转化为具有块结构的单测量矢量模型SMV(Single Measurement Vectors),这样做的目的是便于将目标图像的行间结构先验信息转换为块结构信息,并将此块结构信息用到重构算法里。
由上述可知,构成堆叠矩阵X
d的多个测量向量
具有相同的支撑集,堆叠矩阵X
d中有多少个非零行意味着向量x就有多少个非零块,经过上述转化得到的向量x必定为一个具有块结构特征的稀疏信号,且其中每一块长度都为L。
p(x
i)=N(0,γ
iB
i),i=1,…,SN (12)
其中,γ
i是超参数,其值非负,用来决定第i个块的值是否均为零,B
i则是一个正定矩阵,其值暂 不知,用来对第i个块里各个元素间的结构特征进行建模。
块稀疏信号x的先验为:
p(x)=N(0,∑) (13)
其中:
设公式(11)中,噪声向量v里每个元素均服从高斯分布p(v
i)~N(0,λ)。此时对该红外小目标图像的重构采用块稀疏贝叶斯学习(BSBL)方法,得到:
X
d=ГΦ
T(λI+ΦГΦ
T)
-1Y (15)
其中Γ=diag(γ
1,γ
2,…,γ
SN)。
对各参数的估计如下:
其中,用正定矩阵B作为每一个B
i的最终估计值,目的是为防止过拟合。
针对红外小目标图像的压缩问题,本发明实施例提出一种结合结构信息的红外图像快速重构方法。首先根据红外小目标图像中目标的先验信息,对该红外小目标图像进行行列重整,使得重整后的图像阵列形式满足MMV模型;其次,利用MMV模型对具有相同支撑集的信号进行建模,从而充分利用单帧图像中的行间结构先验信息;最后,将描述时域相关信息的MMV模型转化为具有行间块状相关信息的SMV模型,利用行间结构先验信息的单帧稀疏图像信号恢复红外小目标图像。为验证本实施例提供的重构方法有效性,对采用多种类型的红外图像对算法进行了验证,并与块稀疏贝叶斯学习算法(BSBL)进行了比较。从实验结果看,本实施例提出方法在算法运行时间和重构精度上,均较BSBL算法有提高。
具体实验结果如下:
为验证R-BCS方法的有效性,实验中,本实施例以单目标红外图像和多目标红外图像分别作为实验对象,对重构算法的峰值信噪比和算法重构时间两个技术指标进行了验证。实验中所用图像如图3和图4所示。
考虑到实际中图像编码端通常会对图像的前景图像和背景图像分别进行编码并传输,图像解码端在收到数据后会分别解码前景图像和背景图像,前景和背景经过合成将还原原始图像。基于此,根据本发 明实施例提供的重构方法在实验中首先采用形态学检测方法(top-hat)对该图像进行滤波处理,将目标图像与背景图像相分离,所得到的目标图像如5和图6所示。
图5中,图像像素为320×256,其中目标大小为8×8。该图像具有典型的空域稀疏性。对该红外小目标图像进行下8行采样。采用本发明实施例提供的重构方法对其进行压缩后再实现恢复,所恢复图像的峰值信噪比为106.6191,与原始图像的相对误差为0.1009。对于多目标红外图像图6,多个目标的像素尺寸不太一致,大致为10×7,9×6,和10×6。每个目标的像素尺寸均接近8,此处假设小目标所占像素宽度为8,在实验中对图像进行下8行采样。采用本发明实施例提供的重构方法对其进行压缩后再实现恢复,所恢复图像的峰值信噪比为102.1559,与原图的相对误差为0.1337。
为便于算法的横向比较,分别采用本发明实施例提供的重构方法和BSBL算法对上述图像进行了实验。BSBL算法是20次蒙特卡洛实验,具体实验结果如下表所示:
表1 单目标红外图像重构峰值信噪比比较
对表1和表2中单目标红外图像的重构数据进行比较分析可知,当压缩比不同时,两种方法的各自的重构精度变化不大。当压缩比不变时,R-BCS方法的重构峰值信噪比和BSBL算法不相上下,但是R-BCS方法的重构时间明显远远小于BSBL算法的重构时间。
两种方法在不同压缩比下对多目标小图像的重构峰值信噪比如表3所示,重构时间如表4所示。
表3 多目标红外图像重构峰值信噪比比较
表4 多目标红外图像重构时间比较
对表3和表4中多目标小图像的重构数据进行比较分析可知,当压缩比不同时,两种方法的各自的重构精度变化不大。当压缩比不变时,BSBL算法的重构峰值信噪比比R-BCS方法稍微高一些,但是R-BCS方法的重构时间明显远远小于BSBL算法的重构时间。
本发明实施例根据红外小目标图像的成像特点,结合贝叶斯压缩感知方法,提出了一种结合小目标图像行间结构信息的快速重构方法,即R-BCS方法。该方法可以较为精确地重构单帧目标图像,其还原后的图像平均峰值信噪比较高,快速重构方法采用MMV模型对行间相关性进行建模,降低了重构过程的计算量,提高了算法运行速度。实验结果表明,使用R-BCS方法对红外小目标图像进行压缩,再基于贝叶斯方法对图像进行重构,可以较高精度实现红外小目标图像的还原,且重构方法远远快于BSBL算法。
图7示出了本发明实施例提供的一种基于结构信息的红外小目标图像的快速重构系统,包括:
先验获取单元701,用于获取红外小目标图像中的目标,根据所述目标确定尺寸先验信息;
图像采样单元702,用于根据所述尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;
行堆叠单元703,用于对每一所述子图像进行行堆叠,得到堆叠矩阵,所述堆叠矩阵满足多测量向量模型;
压缩测量单元704,用于对每一所述子图像进行压缩测量,得到包括所述堆叠矩阵的压缩观测矩阵;
图像重构单元705,用于采用块稀疏贝叶斯学习方法对所述压缩观测矩阵进行重构,得到重构后的红外小目标图像。
进一步地,行堆叠单元703具体用于:
进一步地,以X
i·表示所述红外小目标图像X第i行元素所构成的行向量,压缩测量单元704具体用于执行以下步骤:
进一步地,图像重构单元705具体用于执行以下步骤:
Y=Φ·X
d (25)
步骤E2,考虑未知的噪声向量V,将公式(25)进行转换,得到:
Y=Φ·X
d+V (26)
步骤E3,将公式(26)转换成块稀疏模型形式,得到:
y=D·x+v (27)
设定每一个块x
i的概率密度服从高斯分布,即:
p(x
i)=N(0,γ
iB
i),i=1,…,SN; (28)
其中,γ
i表示超参数,用来决定第i个块的值是否均为零,B
i表示正定矩阵,用来对第i个块里各个元素间的结构特征进行建模;
步骤E5,块稀疏信号x的先验为:
p(x)=N(0,∑); (29)
设公式(27)y=D·x+v中的噪声向量v中每个元素服从高斯分布p(v
i)~N(0,λ);
采用块稀疏贝叶斯学习方法对y=D·x+v进行重构,得到重构后的红外小目标图像:
X
d=ГΦ
T(λI+ΦГΦ
T)
-1Y; (30)
其中:Γ=diag(γ
1,γ
2,…,γ
SN);
正定矩阵B作为每一个B
i的最终估计值:
以上所述仅为本发明的较佳实施例而已,并不用以限制本发明,凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等,均应包含在本发明的保护范围之内。
Claims (10)
- 一种基于结构信息的红外小目标图像的快速重构方法,其特征在于,包括:步骤A,获取红外小目标图像中的目标,根据所述目标确定尺寸先验信息;步骤B,根据所述尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;步骤C,对每一所述子图像进行行堆叠,得到堆叠矩阵,所述堆叠矩阵满足多测量向量模型;步骤D,对每一所述子图像进行压缩测量,得到包括所述堆叠矩阵的压缩观测矩阵;步骤E,采用块稀疏贝叶斯学习方法对所述压缩观测矩阵进行重构,得到重构后的红外小目标图像。
- 如权利要求3所述的快速重构方法,其特征在于,以X i·表示所述红外小目标图像X第i行元素所构成的行向量,所述步骤D具体包括:
- 如权利要求4所述的快速重构方法,其特征在于,所述步骤E具体包括:Y=Φ·X d (7)步骤E2,考虑未知的噪声向量V,将公式(7)进行转换,得到:Y=Φ·X d+V (8)步骤E3,将公式(8)转换成块稀疏模型形式,得到:y=D·x+v (9)设定每一个块x i的概率密度服从高斯分布,即:p(x i)=N(0,γ iB i),i=1,…,SN; (10)其中,γ i表示超参数,用来决定第i个块的值是否均为零,B i表示正定矩阵,用来对第i个块里各个元素间的结构特征进行建模;步骤E5,块稀疏信号x的先验为:p(x)=N(0,∑); (11)设公式(9)y=D·x+v中的噪声向量v中每个元素服从高斯分布p(v i)~N(0,λ);采用块稀疏贝叶斯学习方法对y=D·x+v进行重构,得到重构后的红外小目标图像:X d=ΓΦ T(λI+ΦΓΦ T) -1Y; (12)其中:Γ=diag(γ 1,γ 1,…,γ SN);正定矩阵B作为每一个B i的最终估计值:
- 一种基于结构信息的红外小目标图像的快速重构系统,其特征在于,包括:先验获取单元,用于获取红外小目标图像中的目标,根据所述目标确定尺寸先验信息;图像采样单元,用于根据所述尺寸先验信息对所述红外小目标图像进行采样,得到若干子图像;行堆叠单元,用于对每一所述子图像进行行堆叠,得到堆叠矩阵,所述堆叠矩阵满足多测量向量模型;压缩测量单元,用于对每一所述子图像进行压缩测量,得到包括所述堆叠矩阵的压缩观测矩阵;图像重构单元,用于采用块稀疏贝叶斯学习方法对所述压缩观测矩阵进行重构,得到重构后的红外小目标图像。
- 如权利要求8所述的快速重构系统,其特征在于,以X i·表示所述红外小目标图像X第i行元素所构成的行向量,所述压缩测量单元具体用于执行以下步骤:
- 如权利要求9所述的快速重构系统,其特征在于,所述图像重构单元具体用于执行以下步骤:Y=Φ·X d (22)步骤E2,考虑未知的噪声向量V,将公式(22)进行转换,得到:Y=Φ·X d+V (23)步骤E3,将公式(23)转换成块稀疏模型形式,得到:y=D·x+v (24)设定每一个块x i的概率密度服从高斯分布,即:p(x i)=N(0,γ iB i),i=1,…,SN; (25)其中,γ i表示超参数,用来决定第i个块的值是否均为零,B i表示正定矩阵,用来对第i个块里各个元素间的结构特征进行建模;步骤E5,块稀疏信号x的先验为:p(x)=N(0,∑); (26)设公式(24)y=D·x+v中的噪声向量v中每个元素服从高斯分布p(v i)~N(0,λ);采用块稀疏贝叶斯学习方法对y=D·x+v进行重构,得到重构后的红外小目标图像:X d=ΓΦ T(λI+ΦΓΦ T) -1Y; (27)其中:Γ=diag(γ 1,γ 2,…,γ SN);正定矩阵B作为每一个B i的最终估计值:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2018/074506 WO2019148309A1 (zh) | 2018-01-30 | 2018-01-30 | 基于结构信息的红外小目标图像的快速重构方法及系统 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2018/074506 WO2019148309A1 (zh) | 2018-01-30 | 2018-01-30 | 基于结构信息的红外小目标图像的快速重构方法及系统 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019148309A1 true WO2019148309A1 (zh) | 2019-08-08 |
Family
ID=67477859
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2018/074506 WO2019148309A1 (zh) | 2018-01-30 | 2018-01-30 | 基于结构信息的红外小目标图像的快速重构方法及系统 |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO2019148309A1 (zh) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101908889A (zh) * | 2010-07-30 | 2010-12-08 | 哈尔滨工业大学 | 一种块稀疏度未知的稀疏信号压缩感知重构方法 |
US20120259590A1 (en) * | 2011-04-11 | 2012-10-11 | Jong Chul Ye | Method and apparatus for compressed sensing with joint sparsity |
CN104901704A (zh) * | 2015-06-04 | 2015-09-09 | 中国科学院苏州生物医学工程技术研究所 | 具备时空相关特征的躯体传感网信号重构方法 |
CN106663316A (zh) * | 2016-08-30 | 2017-05-10 | 深圳大学 | 一种基于块稀疏压缩感知的红外图像重构方法及其系统 |
CN107147397A (zh) * | 2017-04-24 | 2017-09-08 | 电子科技大学 | 面向可穿戴设备的快速压缩感知重构方法 |
-
2018
- 2018-01-30 WO PCT/CN2018/074506 patent/WO2019148309A1/zh active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101908889A (zh) * | 2010-07-30 | 2010-12-08 | 哈尔滨工业大学 | 一种块稀疏度未知的稀疏信号压缩感知重构方法 |
US20120259590A1 (en) * | 2011-04-11 | 2012-10-11 | Jong Chul Ye | Method and apparatus for compressed sensing with joint sparsity |
CN104901704A (zh) * | 2015-06-04 | 2015-09-09 | 中国科学院苏州生物医学工程技术研究所 | 具备时空相关特征的躯体传感网信号重构方法 |
CN106663316A (zh) * | 2016-08-30 | 2017-05-10 | 深圳大学 | 一种基于块稀疏压缩感知的红外图像重构方法及其系统 |
CN107147397A (zh) * | 2017-04-24 | 2017-09-08 | 电子科技大学 | 面向可穿戴设备的快速压缩感知重构方法 |
Non-Patent Citations (2)
Title |
---|
LIANG, RUNQING: "Structured Compressive Sensing Method and Implementation of Infrared Small Target Image", CHINA MASTER S THESES FULL-TEXT DATABASE (ELECTRONIC JOURNALS), 15 July 2017 (2017-07-15), ISSN: 1674-0246 * |
WANG, FAZONG ET AL.: "Multiple Measurement Vectors for Compressed Sensing: Model and Algorithms Analysis", JOURNAL OF SIGNAL PROCESSING, vol. 28, no. 6, 30 June 2012 (2012-06-30), ISSN: 1003-0530 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103036573B (zh) | 基于滤波的压缩感知信号采集方法 | |
Xiao et al. | Alternating algorithms for total variation image reconstruction from random projections | |
CN106952317B (zh) | 基于结构稀疏的高光谱图像重建方法 | |
CN106663316A (zh) | 一种基于块稀疏压缩感知的红外图像重构方法及其系统 | |
CN107330946A (zh) | 一种基于压缩感知的图像处理方法及装置 | |
CN107192878A (zh) | 一种基于压缩感知的电力系统谐波检测方法及装置 | |
WO2018027584A1 (zh) | 一种目标属性辅助的压缩感知图像恢复方法及其系统 | |
CN103871058B (zh) | 基于压缩采样矩阵分解的红外小目标检测方法 | |
CN102882530B (zh) | 一种压缩感知信号重构方法 | |
Hong et al. | An efficient algorithm for designing projection matrix in compressive sensing based on alternating optimization | |
CN106651974B (zh) | 利用加权结构组稀疏规则的图像压缩感知重构系统及方法 | |
WO2021068495A1 (zh) | 基于平面互质阵列块采样张量信号构造的自由度增强型空间谱估计方法 | |
WO2022110867A1 (zh) | 一种图像压缩采样方法及组件 | |
CN105050105B (zh) | 基于压缩感知的高能效低信息密度数据收集方法 | |
Cui et al. | Deep neural network based sparse measurement matrix for image compressed sensing | |
CN109756740A (zh) | 基于最优测量矩阵的半张量图像压缩方法和图像恢复方法 | |
CN110311685B (zh) | 时序贝叶斯压缩采样及信号解压缩重构方法及数据丢失恢复方法 | |
CN108288295A (zh) | 基于结构信息的红外小目标图像的快速重构方法及系统 | |
Bai et al. | Image fusion and recognition based on compressed sensing theory | |
CN113222860B (zh) | 基于噪声结构多重正则化的图像恢复方法及系统 | |
CN105242237B (zh) | 一种基于压缩感知的电磁矢量阵列参数估计方法 | |
CN113222812A (zh) | 一种基于信息流加强深度展开网络的图像重建方法 | |
WO2019148309A1 (zh) | 基于结构信息的红外小目标图像的快速重构方法及系统 | |
CN104243986B (zh) | 基于数据驱动张量子空间的压缩视频采集与重构系统 | |
CN114624646B (zh) | 一种基于模型驱动复数神经网络的doa估计方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18904381 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 13.11.2020) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 18904381 Country of ref document: EP Kind code of ref document: A1 |