CN104166960B - Self-adaptive non-uniform fixed noise removing method based on scene - Google Patents

Abstract

The invention describes a scene-based self-adaptive non-uniform fixed noise removing method, which is characterized by comprising the following steps of: firstly, determining scene initialization; secondly, performing scene self-adaptive iteration; thirdly, updating the scene template; fourth, an output image is generated. The invention is suitable for removing the non-uniform fixed noise. The method can adapt to the change of the scene, does not generate ghost, has high convergence speed, and solves the convergence problem and ghost problem of the existing algorithm.

Description

Self-adaptive non-uniform fixed noise removing method based on scene

Technical Field

The invention relates to the technical field of non-uniform noise removal, in particular to a scene-based self-adaptive non-uniform fixed noise removal method, which is a self-adaptive non-uniform compensation method and is used for removing fixed mode noise on a focal plane array.

Background

In many applications, focal plane arrays are used to capture images. For example, in target tracking systems, infrared images are captured for tracking using uncooled infrared cameras. The focal plane array comprises infrared detection units arranged in a matrix, for example, there are non-refrigeration infrared cameras having 640 × 480 detection units, and each detection unit outputs one pixel. The responsivity of each detector unit is not exactly the same due to manufacturing process, etc., and there is a certain difference between them. For example, given the same incident radiation for each detection unit, some detection units have large or dark output pixels, where the pixel values output by even similar detection units differ and are not exactly equal. This non-uniform response produces fixed pattern noise. Studies have shown that the non-uniformity of focal plane arrays is related to a number of factors, typically: detector operating temperature, imaging spectrum, response characteristics of the detection unit, etc. Fixed pattern noise is significant in the image, severely impacting image quality, and such fixed pattern noise can also impact target recognition, acquisition, and tracking.

The currently used fixed pattern noise compensation algorithms are basically classified into two categories: firstly, based on a calibrated non-uniform background noise removing algorithm; and secondly, a non-uniform background noise removing algorithm based on the scene. Based on calibrated non-uniform background noise removal algorithms, although the algorithms are simple, the method cannot solve the problem of fixed pattern noise drift, and the adaptability of the method is limited. Although the problem of background drift can be overcome to a certain extent, the adaptive non-uniform background noise removal algorithm based on the scene always has the problem of algorithm convergence or scene ghosting, and most of the adaptive non-uniform background noise removal algorithms do not provide an effective initialization method, so that the application of the existing algorithm is limited due to the problems.

Disclosure of Invention

The invention describes a scene-based self-adaptive non-uniform fixed noise removal method which can adapt to the change of a scene, does not generate ghost images and has high convergence speed.

The technical scheme of the invention is as follows: a self-adaptive non-uniform fixed noise removing method based on scenes comprises the following steps: firstly, determining scene initialization; secondly, performing scene self-adaptive iteration; thirdly, updating the scene template; fourth, an output image is generated.

Step one, in the scene initialization step, it is determined that the scene initialization requires that no attention target is in the input image, when the imaging parameters of the image change, the scene initialization needs to be performed again, and the imaging parameters of the image include the integration time, the aperture size, and the gain size.

In the implementation process, different background images can be stored respectively according to different imaging parameters for scene initialization; the method specifically corresponds to different integration time and aperture size, different background images are recorded according to the gain size, and when the method is used, the corresponding recorded images are directly used as scene initial input.

And step one, acquiring a background image under the current imaging parameter state in the scene determining initialization step.

The iteration process in the scene self-adaptive iteration step in the second step is as follows: assuming that x is an image with non-uniform noise and is output as y after being processed by an algorithm, for the pixel (i, j), the relationship is as follows:

wherein n represents an nth frame image,andrespectively the gain and offset correction coefficients of the picture element (i, j),the pixel (i, j) is expected to output for the pixel value of the position (i, j) in the image where the non-uniform noise exists in the nth frameAnd (i, j) adopting the spatial mean value of the similar neighborhoods, wherein the similar neighborhoods are represented by symbols T and mean areas with similar distances and similar contents, and thus the output value is calculated by a hidden layer in the network:

where t represents the total number of pixels in a similar neighborhood.

The error function is as follows:

wherein M represents an image update template, which is also a neighborhood computed in the iterative process;

relate to itAndthe results are as follows:

step three, in the scene template updating step, the template definition is as follows:

where α represents a threshold value.

The image output process described in step four is as follows:

wherein,indicates the n-th frame correction processing output value,representing the raw data of a non-uniform image,and correcting the output value after data processing.

The invention has the beneficial effects that:

the invention describes a scene-based self-adaptive non-uniform fixed noise removal method which can adapt to the change of a scene, does not generate ghost images and has high convergence speed.

Drawings

FIG. 1 depicts a flow chart of a scene-based adaptive non-uniform fixed noise removal method.

Detailed Description

The invention describes a scene-based self-adaptive non-uniform fixed noise removal method which can adapt to the change of a scene, does not generate ghost images and has high algorithm convergence speed. As shown in fig. 1, a flow chart of the method is depicted. The following describes a specific embodiment of the adaptive non-uniform fixed noise removal method, as follows:

(1) and (5) initializing a scene. The scene initialization in the present invention requires no attention target in the input image, which represents the start of the algorithm in the scene-based adaptive non-uniform fixed noise removal method. When imaging parameters (integration time, aperture size, gain size, etc.) change, scene initialization needs to be performed anew. In implementation, different background images may be stored for different imaging parameters, respectively, for algorithm initialization. For example: the method is characterized in that different background images are respectively recorded corresponding to different integration time, aperture size and gain size, and when the method is used, the corresponding recorded images are directly used as algorithm input. The background image may also be acquired under current imaging parameter conditions.

(2) And (4) self-adaptive iteration. The iterative algorithm of the present invention is based on a neural network correction algorithm based on the assumption that spatially, the signal in the same region is gradual or flat, which is reasonable. The method can overcome the correction error caused by the response drift of the detector to a certain extent, does not require or only needs simple calibration, and adaptively updates the correction coefficient according to the scene information. The iterative method simulates the mechanism of low-layer processing in human retina, assumes the average response of a certain unit neighborhood as the ideal output of the unit, and feeds the value back to the correction coefficient adjusting link, the gain and offset correction coefficient are continuously updated by the adjusting link through the steepest descent method, and then the neuron realizes the self-adaptive correction of each unit. The algorithm process is as follows: assuming that x is an image with non-uniform noise, the image is processed by the algorithm correction unit and then output as y, and for the pixel (i, j), the relationship is as follows:

wherein n represents an nth frame image,andrespectively the gain and offset correction coefficients for the pixel (i, j).The pixel (i, j) is expected to output for the pixel value of the position (i, j) in the image where the non-uniform noise exists in the nth frame(i, j) the spatial mean of the similar neighborhoods, denoted by the symbol T, meaning the areas of similar content and close distances, is used, so that the output value is calculated by the hidden layer in the networkTo:

where t represents the total number of pixels in a similar neighborhood.

The error function is as follows:

wherein M represents an image update template, which is also a neighborhood computed in the iterative process;

relate to itAndthe partial derivatives of (a) are as follows:

and searching the lowest point of the curved surface downwards along the steepest direction of the error performance curved surface by using a steepest descent method. Thereby obtainingAndwhere the iteration step size is mu.

(3) And updating the scene template. In the iterative process, the template needs to be updated, and is defined by a threshold value:

where α represents a threshold value.

(4) And (5) outputting the image. And outputting an image by an algorithm after self-adaptive iteration processing and scene template updating processing. The image output process is as follows:

wherein,indicates the n-th frame correction processing output value,representing the raw data of a non-uniform image,and correcting the output value after data processing.

Claims (6)

1. A self-adaptive non-uniform fixed noise removing method based on scenes is characterized by comprising the following steps:

firstly, determining scene initialization;

secondly, performing adaptive iteration on the scene;

the iteration process in the scene adaptive iteration step described in step two is as follows: assuming that x is an image with non-uniform noise and is output as y after being processed by an algorithm, for the pixel (i, j), the relationship is as follows:

$y_{i,j}^n=a_{i,j}^n*x_{i,j}^n+b_{i,j}^n$

wherein n represents an nth frame image,andrespectively the gain and offset correction coefficients of the picture element (i, j),the pixel (i, j) is expected to output for the pixel value of the position (i, j) in the image where the non-uniform noise exists in the nth frameAnd (i, j) adopting the spatial mean value of the similar neighborhoods, wherein the similar neighborhoods are represented by symbols T and mean areas with similar distances and similar contents, and thus the output value is calculated by a hidden layer in the network:

$f_{i,j}^n=\frac{1}{t}{\mathrm{\Σ}}_{(i,j)\∈T}{y}_{i,j}^n$

wherein t represents the total number of pixels in a similar neighborhood;

the error function is as follows:

$e_{ij}^n={\mathrm{\Σ}}_{(i,j)\∈M}{(y_{i,j}^n-f_{i,j}^n)}^2*M_{i,j}={\mathrm{\Σ}}_{(i,j)\∈M}{(a_{i,j}^n{x}_{i,j}^n+b_{i,j}^n-f_{i,j}^n)}^2*M_{i,j}$

wherein M represents an image update template, which is also a neighborhood computed in the iterative process;

relate to itAndthe results are as follows:

$\frac{\∂e_{i,j}^n}{\∂a_{i,j}^n}=2*{\mathrm{\Σ}}_{(i,j)\∈M}{x}_{i,j}^n*(y_{i,j}^n-f_{i,j}^n)$

$\frac{\∂e_{i,j}^n}{\∂b_{i,j}^n}=2*{\mathrm{\Σ}}_{(i,j)\∈M}(y_{i,j}^n-f_{i,j}^n)$

step three, updating the scene template;

and step four, generating an output image.

2. The method as claimed in claim 1, wherein the second step determines that the scene initialization in the scene initialization step requires no object of interest in the input image, and the scene initialization needs to be performed again when the imaging parameters of the image change, the imaging parameters of the image including integration time, aperture size, and gain size.

3. The method as claimed in claim 1, wherein in the implementation process, different background images can be stored for different imaging parameters respectively for algorithm scene initialization; the specific operation is that different integration time and aperture size are respectively stored, different background images are respectively recorded according to the gain size, and when the method is used, the corresponding recorded images are directly used as the initial value of the algorithm scene to be input.

4. The method of claim 1, wherein the step two is performed to determine that the background image is obtained in the current imaging parameter state in the scene initialization step.

5. The method as claimed in claim 1, wherein the template definition in the scene template updating step in step three is:

where α represents a threshold value.

6. The method for removing the adaptive non-uniform fixed noise based on the scene as claimed in claim 1, wherein the step four generates the output image, and the image output process is as follows:

$y_{i,j}^{n+1}=\left\{\begin{array}{cc}{x}_{i,j}^n,& (i,j)\∈M\\ y_{i,j}^n,& (i,j)\∉M\end{array}\right.$

wherein,indicates the n-th frame correction processing output value,representing the raw data of a non-uniform image,and correcting the output value after data processing.