CN112945897B - Continuous terahertz image non-uniformity correction method - Google Patents

Abstract

The invention discloses a continuous terahertz image non-uniformity correction method, which relates to the technical field of terahertz imaging. The method regards the terahertz intensity signal generated by the continuous terahertz detector pixel as a time series signal, and utilizes the LSTM network to analyze the time series The prediction ability of the signal, according to the part of the terahertz intensity signal generated by the pixel detection, predicts the next terahertz intensity signal, the prediction data is the non-uniform background data corresponding to the current terahertz image, and finally uses the prediction data to rotate and translate the current image Formula non-uniformity correction. The invention does not need to add additional hardware equipment, interrupt the normal operation of the detector, and repeat a large number of complicated statistical calculations and related information calculation and analysis; the invention utilizes the prediction function of the LSTM network for time series signals, and overcomes the impact caused by environmental fluctuations. Influence, generalization ability is strong, and the signal-to-noise ratio of continuous terahertz images is improved.

Description

A method for non-uniformity correction of continuous terahertz images

Technical Field

The invention relates to the technical field of terahertz imaging, and in particular to a method for correcting non-uniformity of continuous terahertz images.

Background Art

Terahertz imaging technology and its application is one of the important research directions of terahertz technology, and has broad application prospects in non-destructive testing, security inspection, food and drug safety testing, medical and biological imaging, etc. The imaging quality of terahertz wave imaging system depends on the development of terahertz source, terahertz detector and imaging technology. Although the image quality has been partially improved by improving the hardware of terahertz wave imaging system, it is limited by factors such as terahertz detector manufacturing process, imaging optical path and working environment, resulting in uneven response between different pixels, and even a series of problems such as blind pixels. Therefore, it is necessary to perform non-uniformity correction processing on terahertz images.

At present, the main methods for terahertz image non-uniformity correction are calibration-based methods and scene-based methods. Calibration-based methods include one-point calibration, two-point calibration and multi-point calibration. However, all three methods require periodic interruption of the detector and irradiation with a specific power terahertz source for calibration correction. The scene-based method uses the collected image to estimate the correction coefficient of the detector, such as the time domain high-pass filter method, the neural network correction method, etc. The scene-based method uses the imaging information in multiple scenes and can effectively estimate the correction parameters in the actual scene without recalibration. However, when the scene changes, fixed pattern noise will remain, affecting the imaging quality.

In order to overcome the above technical problems, it is urgent to propose a new correction method to perform non-uniformity correction processing on the non-uniformity problem of continuous terahertz detectors, so as to better achieve terahertz imaging effects.

Summary of the invention

In view of this, it is necessary to provide a method for correcting the non-uniformity of continuous terahertz images to address the above-mentioned problems. The terahertz intensity signal generated by the continuous terahertz detector pixels is regarded as a time series signal. The prediction ability of the LSTM network for time series signals is used to predict the next terahertz intensity signal based on the partial terahertz intensity signal generated by the pixel detection. The predicted data is the non-uniform background data corresponding to the current terahertz image. Finally, the predicted data is used to perform rotational and translational non-uniformity correction on the current image.

To achieve the above object, the present invention is implemented according to the following technical solutions:

A method for correcting non-uniformity of continuous terahertz images comprises the following steps:

Step S1: using a continuous terahertz imaging system to collect non-uniform terahertz intensity images without a detection object, constructing a background image sample set y; collecting terahertz intensity images when the terahertz source is turned off, forming an image sample set z, the number of sample sets z is M; the image sizes in sets y and z are both m×n, where m is the number of continuous terahertz detector imaging units, i.e., the number of rows of a single image, and n is the number of columns of a single image;

Step S2: constructing an LSTM network model for each imaging unit of the continuous terahertz detector, with a total of m LSTM network models;

Step S3: Taking the first n ₁ intensity data of the i-th row of each image in the background sample image set y as input and the n ₁ intensity data after the i-th row as expected values, the i-th LSTM network model is trained; repeat the above steps until the LSTM network model training corresponding to the m pixels of the continuous terahertz detector is completed;

Step S4: for the image x to be non-uniformity corrected with a size of m×n, take the first n ₁ intensity data of the i-th row as the input of the trained i-th LSTM network model, and predict and generate n ₁ intensity data;

Step S5: Repeat step S4 for each row of image x to generate an m×(nn ₁ ) predicted intensity image, and merge the m×n ₁ image before image x and the predicted m×(nn ₁ ) predicted image into an m×n image.

第i行进行均值计算，得到均值

Step S6: Image

The mean is calculated for the i-th row, and the mean is obtained.

计算公式为：Step S7: For the image sample set z, calculate the bias parameter of the i-th imaging unit of the terahertz detector

The calculation formula is:

Where z _kij is the intensity value of the i-th row and j-th column of the k-th image;

Step S8: Perform non-uniformity correction on each pixel _xij of the image x, and the correction formula is:

为像素值x_ij校正后的值，各像素校正后构成校正后图像

in,

is the corrected value of the pixel value x _ij , and each pixel after correction constitutes the corrected image

Compared with the prior art, the advantages and positive effects of the present invention are:

Compared with the non-uniformity correction method based on calibration, the method proposed by the present invention does not need to add additional hardware equipment and does not need to interrupt normal work; compared with the non-uniformity correction method based on scene, it does not need to repeat a large number of complex statistical calculations and related information calculation and analysis.

The method proposed in the present invention utilizes the prediction function of the LSTM network for time series signals to predict the next non-uniform background based on the partial pixels of the image to be corrected. The method described in the present invention can generate the non-uniform background corresponding to the current image, overcome the influence of environmental fluctuations, and has strong generalization ability. In addition, the non-uniformity correction method described in the present invention improves the signal-to-noise ratio of continuous terahertz images, provides a reliable image analysis source for the application of continuous terahertz wave images, and expands the application range of terahertz imaging systems.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

FIG1 is a schematic flow chart of a method for correcting non-uniformity of continuous terahertz images according to the present invention.

DETAILED DESCRIPTION

In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be pointed out that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Example 1

As shown in Figure 1, the present invention proposes a dynamic non-uniformity correction method based on a long short-term memory neural (LSTM) network to address the non-uniformity problem of continuous terahertz detectors. The method of the present invention regards the terahertz intensity signal generated by the pixels of the continuous terahertz detector as a time series signal, and uses the LSTM network's ability to predict time series signals to predict the next terahertz intensity signal based on the partial terahertz intensity signal generated by the pixel detection. The predicted data is the non-uniform background data corresponding to the current terahertz image, and finally the predicted data is used to perform rotational translational non-uniformity correction on the current image.

该图像为待校正图像x对应的非均匀性背景图像。再计算图像

第i行强度数据的均值

作为成像单元的增益校正系数。再对图像样本集x中强度图像按行进行均值计算，均值结果

作为第i个成像单元的偏置校正系数。最后利用校正公式对待校正强度图像x中像素x_ij进行校正。The technical solution of the present invention first uses a continuous terahertz imaging system to collect non-uniform terahertz intensity images in the absence of a detection object, and constructs an image sample set y; the terahertz intensity image when the terahertz source is turned off constitutes an image sample set z, and the image sizes in sets y and z are both m×n. Then, a long short-term memory neural network model is constructed for each imaging unit of the terahertz detector, and it is trained using the image set y. During the training process, the first n1 intensity data of the i-th row of an intensity image in y are taken as the input of the i-th LSTM network, and the last n-n1 intensity data are the expected output. Then, the first n1 columns of data of the intensity image x to be corrected are input into the trained m LSTM networks, and n-n1 columns of terahertz intensity values are predicted and generated, and the predicted intensity data and the first n1 columns of intensity data are merged into an image of size m×n.

This image is the non-uniform background image corresponding to the image to be corrected x. Then calculate the image

The mean of the intensity data in the i-th row

As the gain correction coefficient of the imaging unit. Then the intensity image in the image sample set x is averaged by row, and the average result is

As the bias correction coefficient of the i-th imaging unit. Finally, the correction formula is used to correct the pixel _xij in the intensity image x to be corrected.

Specifically, the continuous terahertz image non-uniformity correction method of the present invention comprises the following steps:

Step S1: using a continuous terahertz imaging system to collect non-uniform terahertz intensity images without a detection object, constructing a background image sample set y; collecting terahertz intensity images when the terahertz source is turned off, forming an image sample set z, the number of sample sets z is M; the image sizes in sets y and z are both m×n, where m is the number of imaging units of the continuous terahertz detector, that is, the number of rows of a single image, and n is the number of columns of a single image;

Step S2: constructing an LSTM network model for each imaging unit of the continuous terahertz detector, with a total of m LSTM network models;

Step S3: Taking the first n ₁ intensity data of the i-th row of each image in the background sample image set y as input and the n ₁ intensity data after the i-th row as expected values, the i-th LSTM network model is trained; repeat the above steps until the LSTM network model training corresponding to the m pixels of the continuous terahertz detector is completed;

Step S4: for the image x to be non-uniformity corrected with a size of m×n, take the first n ₁ intensity data of the i-th row as the input of the trained i-th LSTM network model, and predict and generate n ₁ intensity data;

Step S5: Repeat step S4 for each row of image x to generate an m×(nn ₁ ) predicted intensity image, and merge the m×n ₁ image before image x and the predicted m×(nn ₁ ) predicted image into an m×n image.

第i行进行均值计算，得到均值

Step S6: Image

The mean is calculated for the i-th row, and the mean is obtained.

计算公式为：Step S7: For the image sample set z, calculate the bias parameter of the i-th imaging unit of the terahertz detector

The calculation formula is:

Where z _kij is the intensity value of the i-th row and j-th column of the k-th image;

Step S8: Perform non-uniformity correction on each pixel _xij of the image x, and the correction formula is:

为像素值x_ij校正后的值，各像素校正后构成校正后图像

in,

is the corrected value of the pixel value x _ij , and each pixel after correction constitutes the corrected image

Compared with the prior art, the method proposed in the present invention does not require the addition of additional hardware equipment or interruption of normal operation compared with the calibration-based non-uniformity correction method; and does not require the repetition of a large number of complex statistical calculations and related information calculation and analysis compared with the scene-based non-uniformity correction method.

The method proposed in the present invention utilizes the prediction function of the LSTM network for time series signals to predict the next non-uniform background based on the partial pixels of the image to be corrected. The method described in the present invention can generate the non-uniform background corresponding to the current image, overcome the influence of environmental fluctuations, and has strong generalization ability. In addition, the non-uniformity correction method described in the present invention improves the signal-to-noise ratio of continuous terahertz images, provides a reliable image analysis source for the application of continuous terahertz wave images, and expands the application range of terahertz imaging systems.

The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims (1)

1. A continuous terahertz image non-uniformity correction method is characterized by comprising the following steps:

step S1: acquiring a non-uniform terahertz intensity image under the condition of no detection object by using a continuous terahertz imaging system, and constructing a background image sample set y; collecting terahertz intensity images when a terahertz source is turned off to form an image sample set z, wherein the number of the sample set z is M; the sizes of the images in the sets y and z are both mxn, wherein m is the number of imaging units of the continuous terahertz detector, namely the number of rows of a single image, and n is the number of columns of the single image;

step S2: constructing an LSTM network model for each imaging unit of the continuous terahertz detector, wherein m LSTM network models are constructed;

and step S3: using background sample image to set y image ith row n before ₁ The intensity data is input, n-n after the ith line ₁ Training the ith LSTM network model with the intensity data as expected values; repeating the steps until the LSTM network model training corresponding to the m imaging units of the continuous terahertz detector is completed;

and step S4: taking the image x to be non-uniformity corrected with the size of m multiplied by n, and taking the n before the ith row ₁ Using the intensity data as the input of the ith LSTM network model after training to predict and generate n-n ₁ (ii) intensity data;

step S5: repeating step S4 for each line of image x to generate mx (n-n) ₁ ) Predicting the intensity image, and moving the image x front by m × n ₁ Image and prediction generated mx (n-n) ₁ ) Merging of predicted pictures into an mxn sized picture

Step S6: for images

The ith row carries out the average value calculation to obtain the average value->

Step S7: calculating the bias parameter of the ith imaging unit of the terahertz detector for the image sample set z

The calculation formula is as follows:

wherein z is _kij The ith row and the jth column of the kth image are used as the intensity value;

step S8: for each pixel x of the image x _ij And carrying out non-uniformity correction, wherein the correction formula is as follows:

wherein,

is a pixel value x _ij The corrected value, corrected for each pixel, constitutes a corrected image->

Images