Image processing method for filtering IRFPA vertical stripe noise based on transverse FIR technology

Abstract

The invention provides an image processing method for filtering IRFPA vertical stripe noise based on a transverse FIR (Finite Impulse Response) technology, which belongs to the technical field of infrared image noise filtering and comprises the following steps of: acquiring an original image, setting parameters of a filter according to the size of the image, solving an average value of each column of a video, forming a column average row data, and obtaining a row average row data; the process is called column averaging or longitudinal averaging; performing high-pass FIR filtering on the column-direction average data to obtain an interference signal; local adjustment in the row direction; local adjustment in the column direction; the filtered image is obtained by subtracting the interference signal line by line, so that the defects of large calculation amount, loss of image details, locality of an algorithm and the like of a spatial median filtering method are effectively overcome, and meanwhile, the method has flexible spatial and local adjustment capability; and the problem that the frequency domain adjustment method is not flexible and visual is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of infrared image noise filtering, and more specifically to an image processing method for filtering IRFPA vertical stripe noise based on a transverse FIR technology. Background Art

[0002] Banding noise is a special type of image non-uniformity. Possible causes include:

[0003] Pixel response differences will lead to image non-uniformity. A column or a row of detectors share the same amplifier circuit, so column or row stripe non-uniformity will occur. In addition, the bias voltage ripple corresponding to the infrared focal plane pixel will also lead to pixel output non-uniformity. Conventional calibration-based or scene-based non-uniformity correction methods can effectively remove some stripes, but they are ineffective for some strong stripes.

[0004] Inconsistencies caused by multi-channel signal acquisition circuits and system working conditions; each analog signal is processed separately, and the analog signal is processed by amplification, filtering, differentiation, ADC, etc.; when designing the system, a unified signal source input is used for each channel, and the output is kept as consistent as possible by fine-tuning the chip peripheral devices such as amplification, filtering, differentiation, ADC, etc.; due to the long-term operation of the system, the typical values ​​of the devices will shift and cause slight differences between the signals; through the statistics of the strip mean, it can be concluded that the strip code value and the surrounding code value are very small, with only a single-digit difference; the infrared image non-uniformity correction method can better correct fixed pattern noise and regular stripe noise; it does not work for strong stripes in the original image without filtering out the vertical stripe noise, which may be caused by the inconsistency caused by the system working condition; multi-channel processing of analog signals will bring stripe noise in the column direction due to the difference in processing channels.

[0005] Common methods for removing banding noise are:

[0006] Median filter method; Median filter method is a commonly used nonlinear filter; The basic principle is to select the median of each pixel value in a neighborhood of the pixel to be processed to replace the processed pixel. Its main function is that the gray value of the pixel is close to the surrounding pixels, so as to eliminate isolated noise points, so the median filter can well eliminate impulse noises such as salt and pepper noise and Gaussian noise. Not only that, while eliminating noise, the median filter can also effectively protect the boundary information of the image and will not cause great blurring to the image (compared to the mean filter); The noise processing effect of the median filter is determined by the processing window size, but the window size has a contradiction between eliminating noise and protecting the details of the image: if the filter window is small, some details in the image can be well protected, but the filtering effect of noise is not very good; On the contrary, increasing the window size has a relatively good noise filtering effect, but it will make the image blurred and lose image detail information; In addition, according to the principle of the median filter, if the number of noise points in the filter window is greater than the number of pixels in the entire window, the median filter cannot filter the noise well, causing the median filter function to fail.

[0007] Using the Fourier transform formula to transform the signal into the frequency domain for noise filtering is one of the basic methods of transform domain denoising. Applying this method to the filtering of stripe noise in hyperspectral images requires first transforming a hyperspectral image containing stripe noise into the Fourier frequency domain space, where each spectrum region corresponds to a specific component in the spatial domain. Based on the analysis of the specific component of stripe noise generated by different spectrometers, a filter can be constructed in the Fourier frequency domain space to filter out the frequency component of stripe noise, and finally the image is obtained by inversely transforming it into the time domain. However, the frequency domain denoising method does not retain detailed information well while filtering out noise, and most of them are computationally intensive and have high algorithm complexity.

[0008] The polynomial fitting method counts the mean and variance of the image row or column direction, and then uses polynomial fitting. Polynomial fitting plays a smoothing role and fits the outlier mean or variance of some rows or columns to the curve. This method can effectively remove stripe noise and is simple. The disadvantage is that the fitting order is uncertain and requires multiple experiments, which is computationally intensive.

[0009] In order to solve these problems, an image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology is provided. Summary of the invention

[0010] The present invention provides an image processing method for filtering out IRFPA vertical stripe noise based on a transverse FIR technology, which can utilize the verticality, slow variation, high frequency and small fluctuation characteristics of stripe noise, effectively overcome the shortcomings of a spatial median filtering method such as a large amount of calculation, loss of image details and locality of an algorithm, and at the same time, compared with a method for removing stripes in a frequency domain, in addition to significantly reducing the amount of calculation, the method has flexible spatial and local adjustment capabilities; and although the frequency domain processing adopts an FFT fast algorithm, the amount of calculation is still very large, and the real part and the imaginary part of the frequency domain are Hilbert transform pairs, and the frequency domain adjustment method is inflexible and unintuitive.

[0011] The image processing method for filtering out IRFPA vertical stripe noise based on the horizontal FIR technology includes the following steps: obtaining the original image, setting the filter parameters according to the image size, calculating the average value of each column of a video, and forming a column average row data, this process is called column average or vertical average; performing high-pass FIR filtering on the column average data to obtain an interference signal; local adjustment in the row direction; local adjustment in the column direction; subtracting the interference signal row by row to obtain a filtered image.

[0012] Furthermore, in step S1, the width of the original image is W and the height is H. The width W of the image is divided into N regions according to the threshold T1, that is, the image is divided into N sub-regions of size W / N*H. After obtaining each region, the mean of the column vector is calculated for each sub-region, and the global mean of the current sub-region is calculated.

[0013] Furthermore, in step S2, the mean vector is set according to the mean calculation value of S1, and the mean vector is expanded in the row direction using a reasonable threshold, and the expansion size is positively correlated with T1, that is:

[0014]

[0015] Where: P = W / T1

[0016] W is the width of the image;

[0017] T1 is the threshold set in S1;

[0018] ln1 represents the first value of the column mean;

[0019] m represents the mth position of the filter;

[0020] This filter sequence exLn is introduced into the FIR filter.

[0021] Furthermore, the numerator coefficient of the FIR filter is determined by the window function fir, the order of fir is equal to P, and the cutoff frequency is set to T2, where T2 is the normalized cutoff frequency and ranges from 0 to 1.

[0022] Furthermore, in step S3, the absolute value of the stripe interference is determined in the row direction, and the absolute value exceeding 0.2°C is determined to be generated by the background, and the interference here is replaced by 0, and then the corrected interference is subtracted from the original image to obtain the filtered image.

[0023] Furthermore, in step S4, the entire image is segmented in the column direction and processed independently.

[0024] Furthermore, in step S5, the separated interference signal is approximately equal to the interference signal of each row, and the interference signal is subtracted from the original image to obtain a relatively pure background image.

[0025] Beneficial effects: By utilizing the verticality, slow variation, high frequency and small fluctuation characteristics of stripe noise, the shortcomings of the spatial median filtering method, such as large amount of calculation, loss of image details and locality of the algorithm, are effectively overcome. At the same time, compared with the method of removing stripes in the frequency domain, in addition to significantly reducing the amount of calculation, it has flexible spatial and local adjustment capabilities; and although the frequency domain processing adopts the FFT fast algorithm, the amount of calculation is still very large, and the real and imaginary parts of the frequency domain are Hilbert transform pairs, and the frequency domain adjustment method is inflexible and unintuitive. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Filtering process flow chart;

[0027] Figure 2 The original image without filtering out vertical stripe noise;

[0028] Figure 3 against Figure 2 The image obtained by taking the average value of each column in;

[0029] Figure 4 The interference signal is obtained by high-pass FIR filtering of the original image. The yellow line in the figure is the interference signal;

[0030] Figure 5 Image after high-pass FIR filtering; DETAILED DESCRIPTION

[0031] like Figure 1 and 2 As shown in the figure, an image processing method for filtering out IRFPA vertical stripe noise based on the horizontal FIR technology includes the following steps: obtaining the original image, setting the parameters of the filter according to the image size, calculating the average value of each column of a video, and forming a column average row data. This process is called column average or vertical average; performing high-pass FIR filtering on the column average data to obtain an interference signal; local adjustment in the row direction; local adjustment in the column direction; subtracting the interference signal row by row to obtain a filtered image.

[0032] Get the original image, set the filter parameters according to the image size, find the average value of each column in a video, and form a column-averaged row data. This process is called column-wise averaging or vertical averaging, that is, accumulation by row.

[0033] Due to stripe interference, the word "interference" is used here instead of noise to emphasize that the changes are very slow and vertical. It can be considered that the interference in each row is the same; so when n rows are accumulated, the interference is enhanced to n times that of a single row, and the background fluctuations in the column direction are averaged, which will cause the background component in the vertical average to be more inclined to the low frequency band; especially for mottled backgrounds, this smoothing effect is more obvious.

[0034] The width of the original image is W and the height is H. The width W of the image is divided into N regions according to the threshold T1, that is, the image is divided into N sub-regions of size W / N*H. After obtaining each region, the mean of the column vector is calculated for each sub-region, and the global mean of the current sub-region is calculated.

[0035] like Figure 3 As shown, the interference signal is obtained by performing high-pass FIR filtering on the column-wise average data.

[0036] After the above processing, it can be seen intuitively that the stripe interference is a high-frequency interference. Due to its vertical characteristics, its frequency characteristics do not change after longitudinal averaging; and the background is pushed to low frequency, which makes the stripe interference that is originally in the high-frequency band easier to separate; the most significant feature of the FIR filter compared to the IIR filter is that it is easy to control the linear phase, that is, the interference signal can be completely separated through the FIR high-pass filter.

[0037] According to the mean calculation value in step 1, the mean vector is set, and the mean vector is expanded in the row direction using a reasonable threshold. The expansion size is positively correlated with T1.

[0038]

[0039] Where P = W / T1, where W is the width of the image, T1 is the threshold set in step 1, ln1 represents the first value of the column mean, and m represents the mth position of the filter; this filter sequence exLn is imported into the FIR filter; wherein the numerator coefficient of the FIR filter is determined by the window function fir; here the order of fir is equal to P, and the cutoff frequency is set to T2, where T2 is the normalized cutoff frequency, ranging from 0 to 1.

[0040] Local adjustment in the row direction; at the longitudinal edge of the object, or some extreme longitudinal striped objects themselves, the grayscale in the frequency domain may be close to the stripe interference; at this time, if FIR filtering is performed without adjustment, it will bring the opposite effect of filtering; actual measurements show that the equivalent temperature difference of the vertical stripes of IRFPA is below 0.2℃, and only when the temperature difference range of the entire image is small, the interference stripes will appear after histogram stretching; in order to make full use of the interference stripes, the absolute value of the stripe interference is determined before executing FIR. For those with an absolute value exceeding 0.2℃, it is determined to be generated by the background, and the interference here is replaced by 0.

[0041] Frequency domain filtering is global in space and it is difficult to make similar local adjustments. When there are columnar (long) objects in the background image, frequency de-striking filtering will produce unacceptable shadows above and below the columnar objects.

[0042] Local adjustment in the column direction: When an image has ups and downs in the upper part and is very flat in the lower part, the stripes in the lower part may not be processed due to step 4. If step 4 is removed, the shadow of the upper part may appear in the lower part.

[0043] If the whole image is divided into two halves and processed separately, a good effect will be obtained. Furthermore, if the whole image is evenly divided into several sections along the vertical direction and processed independently, a good visual effect will be obtained.

[0044] like Figure 4 and 5 As shown, the filtered image is obtained by subtracting the interference signal line by line; the interference signal separated from the previous step is approximately equal to the interference signal in each line; a relatively pure background image can be obtained by subtracting the interference signal from the original image.

[0045] Without considering step 4, the present invention only performs the FIR filtering algorithm for one row of data on the entire image. Column averaging also only needs to traverse the pixels once. Therefore, the algorithm of the present invention has a very small amount of computation. Considering the situation of step 4, although the overall amount of computation increases, each segment is processed completely independently. The algorithm was tested in a multi-core CPU and it was found that when the number of segments is close to the number of CPU cores, the computing time is shortened.

[0046] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology, characterized in that: The following steps are involved: S1, obtain the original image, and set the filter parameters according to the image size, find the average value of each column of a video, and form a column average row data. This process is called column average or vertical average; S2, high-pass FIR filtering is performed on the column-wise average data to obtain the interference signal; S3, local adjustment in row direction; S4, local adjustment in column direction; S5, subtracting the interference signal line by line to obtain a filtered image.

2. The image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology according to claim 1, characterized in that: In step S1, the width of the original image is W and the height is H. The width W of the image is divided into N regions according to the threshold T1, that is, the image is divided into N sub-regions of size W / N*H. After each region is obtained, the mean of the column vector is calculated for each sub-region, and the global mean of the current sub-region is calculated.

3. The image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology according to claim 1, characterized in that: In step S2, the mean vector is set according to the mean calculation value of S1, and the mean vector is expanded in the row direction using a reasonable threshold, and the expansion size is positively correlated with T1, that is: Where: P = WT1 W is the width of the image; T1 is the threshold set in S1; ln1 represents the first value of the column mean; m represents the mth position of the filter; This filter sequence exLn is introduced into the FIR filter.

4. The image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology according to claim 3, characterized in that: The numerator coefficient of the FIR filter is determined by the window function fir. The order of fir is equal to P. The cutoff frequency is set to T2. T2 is the normalized cutoff frequency and ranges from 0 to 1.

5. The image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology according to claim 1, characterized in that: In step S3, the absolute value of the fringe interference is determined in the row direction, and the absolute value exceeding 0.2°C is determined to be generated by the background, and the interference here is replaced by 0, and then the corrected interference is subtracted from the original image to obtain the filtered image.

6. The image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology according to claim 1, characterized in that: In step S4, the entire image is segmented in the column direction and processed independently.

7. The image processing method for filtering out IRFPA vertical stripe noise based on horizontal FIR technology according to claim 1, characterized in that: In step S5, the separated interference signal is approximately equal to the interference signal of each row, and the interference signal is subtracted from the original image to obtain a relatively pure background image.

Images