Small target infrared image processing method based on weighted local image entropy
Technical Field
The invention relates to the technical field of digital image processing, in particular to a small-target infrared image processing method based on weighted local image entropy.
Background
The small target infrared image processing technology is widely applied to the civil field (such as satellite atmospheric infrared cloud picture analysis, infrared medical image pathological analysis, geological analysis, sea surface personnel search and rescue, intrusion detection and forest fire detection) and the military field (such as accurate guidance, early warning detection, battlefield command and reconnaissance and friend-foe identification).
The small target image has small target, weak strength, no prior size, shape, texture and other characteristics, and the target, background and noise are mixed together and difficult to detect directly. However, the background is generally considered to have correlation in the spatial domain, stability in the temporal domain, and be in the low frequency part of the image in the frequency domain, while the target is generally considered to be uncorrelated with the background in the spatial domain and be in the high frequency part of the image in the frequency domain. Therefore, the small target infrared image processing algorithm is mainly divided into three types of time domain, space domain and transform domain: the time domain algorithm is mainly used for inhibiting the background with short-time stationarity, but the inhibiting effect on the complex background is not ideal. The space domain algorithm has good real-time performance and is easy to realize. The median filtering is only suitable for eliminating random noise with the pulse width smaller than a filtering window, and cannot process a structured background; the top-hat transformation is a practical nonlinear background filtering technology, but needs prior knowledge of images, and has poor adaptivity; adaptive filtering techniques such as two-dimensional minimum mean square error filtering and other algorithms require that the statistical characteristics of the background are unchanged or slowly change, so that the complex background cannot be effectively suppressed. The transform domain algorithm is based on adaptive frequency domain Butterworth high-pass filtering, wavelet transform and the like, but the algorithm is derived from Fourier transform and is limited by a Heisenberg (Heisenberg) inaccuracy measuring principle (namely, the product of a time window and a frequency window is a constant), and the algorithm needs to be transformed twice in a positive and negative way, so that the algorithm has a large operation amount.
Although many achievements have been achieved in the field of small target infrared image processing, and many algorithms have been well implemented in engineering applications, the target detection system engineering still faces great difficulty and complexity for small target infrared images with low signal-to-noise ratio under a complex background. How to design a small target infrared image processing algorithm with simple structure, good filtering effect and strong robustness is a key problem of target detection technology research.
Disclosure of Invention
Aiming at the technical problems of the existing small target infrared image processing method, the invention provides a small target infrared image processing method based on weighted local image entropy.
A small target infrared image processing method based on weighted local image entropy comprises the following steps:
step 1, solving the multi-scale gray difference D of each pixel point (x, y) of the image;
step 2, solving the local image entropy E of each pixel point (x, y) of the image;
step 3, obtaining the weighted local image entropy H of each pixel point (x, y) through the multi-scale gray level difference D and the local image entropy E;
and 4, solving an adaptive threshold T according to the weighted local image entropy H, and carrying out binarization on the weighted local image entropy H through the adaptive threshold T to detect the infrared small target.
The multi-scale grayscale difference D of step 1 as described above is solved by:
step 1.1 for infraredThe gray value corresponding to each pixel point (x, y) in the image I is I (x, y), and the maximum neighborhood space omega of the pixel point (x, y) is setmaxThe neighborhood space omegamaxIs of size Lmax×LmaxWherein L ismaxIs a positive odd number greater than 1;
step 1.2, obtaining a neighborhood space set { omega ] of each pixel point (x, y)k1,2, …, L, where L (L) ismax-1)/2,ΩkThe size of (2 · k +1) × (2 · k + 1);
step 1.3, calculating the neighborhood omega of each pixel point (x, y) by using the following formulakAnd omegamaxGray difference D betweenk(x,y),k=1,2,…,L:
Wherein,andrespectively representing the neighborhoods Ωk、ΩmaxThe number of intra-pixel points, I (s, t), represents the neighborhood ΩkThe gray value at the inner point (s, t), I (p, q), represents the neighborhood ΩmaxGray value at inner point (p, q);
step 1.4, calculating the multi-scale gray difference D (x, y) corresponding to each pixel point (x, y):
D(x,y)=max{D1(x,y),D2(x,y),...,DL(x,y)}。
the local image entropy E of step 2 as described above is solved by:
setting a neighborhood space theta of each pixel point (x, y) in the infrared image I, wherein the size of the neighborhood space theta is mxn, and calculating the local image entropy at the pixel point (x, y):
wherein, the constant is a set normal number, I (I, j) represents a gray value at a point (I, j) in the neighborhood Θ, and each pixel point in the infrared image I is traversed to obtain the local image entropy E of the infrared image I.
The weighted local image entropy H of step 3 as described above is solved by:
and (3) performing dot product operation on the multi-scale gray difference D obtained by processing each pixel point (x, y) in the step (1) and the local image entropy E obtained by processing in the step (2) to obtain the weighted local image entropy H corresponding to each pixel point (x, y).
The adaptive threshold T as described above is determined by the following equation:
T=c·SNR·σ+mm,SNR=(Hmax-mm)/σ
where c is a positive constant, σ is the standard deviation of the weighted local image entropy H, mm is the mean of the weighted local image entropy H, HmaxIs the maximum value of the weighted local image entropy H.
Compared with the prior art, the invention has the following advantages:
1. the method utilizes the characteristics of the target and the background in the small target infrared image, does not depend on an infrared image model and parameter selection, can effectively inhibit the background and the noise of the infrared image, and improves the signal-to-noise ratio of the infrared image, thereby improving the detection probability of the target and reducing the false alarm probability.
2. According to the method, a multi-scale gray level difference graph of the infrared image is constructed, so that a large amount of noise interference can be eliminated; secondly, obtaining a weighted local image entropy through dot product operation, wherein the obtained weighted local image entropy graph has high signal-to-noise ratio gain and can effectively inhibit background and noise; and then, the target is detected by using the self-adaptive threshold, so that the problems of unstable image processing, self-adaptability and the like under the complex background condition are solved.
Drawings
FIG. 1 is a block flow diagram of the present invention.
Fig. 2 is a comparison graph of the processing result schematic diagram obtained by the method of the embodiment 1 and the processing result schematic diagram of the prior art algorithm. A is a small target infrared original image of a sea-air background, B is a filtering result adopting a multi-scale gray difference operator, C is a filtering result adopting a local image entropy operator, D is a weighted local image entropy image, and E is a detection result adopting a self-adaptive threshold value.
Fig. 3 is a schematic diagram of an infrared image processing result obtained by the method of the prior art and the present embodiment. (a _1), (B _1), (C _1), (D _ 1): sequentially obtaining low signal-to-noise ratio small target infrared images under different backgrounds and noise degrees; (a _2), (B _2), (C _2), (D _ 2): filtering results of the maximum background prediction model-based method corresponding to (a _1), (B _1), (C _1), and (D _1) in this order; (a _3), (B _3), (C _3), (D _ 3): top-hat operator based filtering results corresponding to (a _1), (B _1), (C _1), (D _1) in order; (a _4), (B _4), (C _4), (D _ 4): the filtering results of step 1 to step 3 of the method of the present embodiment, which correspond to (a _1), (B _1), (C _1), and (D _1) in sequence; (a _5), (B _5), (C _5), (D _ 5): the results of detection of the infrared small target based on the method of the present embodiment sequentially correspond to (a _1), (B _1), (C _1), and (D _ 1).
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.
Example 1:
FIG. 1 shows that the method mainly comprises the following steps: image input, multi-scale gray difference operator solving, local image entropy operator solving, dot product operation, adaptive threshold solving and binaryzation.
The method specifically comprises the following steps:
step 1, inputting an infrared image, and solving the multi-scale gray difference D of the image:
the small target infrared image is generally composed of three parts of a target, a background and noise. The imaged size of a small object is generally less than 80 pixels, i.e. less than 0.12% of 256 × 256, so the object has no features such as size, shape and texture, but it differs from background and noise in terms of gray value, frequency and correlation. The core idea of the multi-scale gray scale difference operator (D) is to utilize the gray scale difference between a target area and a target neighborhood in a small target infrared image, and inhibit the background and enhance the target through the measurement of the difference.
The solving process of the multi-scale gray scale difference operator D of the infrared image I is as follows:
(1) for each pixel point (x, y) in the infrared image I, the corresponding gray value is I (x, y), and the maximum neighborhood space omega of the pixel point (x, y) is setmaxThe neighborhood space omegamaxIs of size Lmax×LmaxWherein L ismaxIs a positive odd number greater than 1;
(2) obtaining a neighborhood space set [ omega ] of each pixel point (x, y)k1,2, …, L, where L (L) ismax-1)/2,ΩkThe size of (2 · k +1) × (2 · k + 1);
(3) calculating the neighborhood omega of each pixel point (x, y)kAnd omegamaxGray difference D betweenk(x,y),k=1,2,…,L:
Wherein,andrespectively representing the neighborhoods Ωk、ΩmaxThe number of intra-pixel points, I (s, t), represents the neighborhood ΩkThe gray value at the inner point (s, t), I (p, q), represents the neighborhood ΩmaxGray value at inner point (p, q).
(4) Calculating the multi-scale gray difference D (x, y) corresponding to each pixel point (x, y):
D(x,y)=max{D1(x,y),D2(x,y),...,DL(x,y)} (2)
and traversing each pixel point in the infrared image I to obtain the multi-scale gray difference D (shown as B in FIG. 2) of the infrared image I. As can be seen from B of fig. 2, the background of the infrared image I is suppressed and the target is well enhanced.
Step 2, solving the local image entropy E of the image:
for the background of the infrared image I, the texture features are determined, when an object appears in the image, the texture features of the image are destroyed, and the small object has a small contribution to the entropy value of the whole image, but in the local window, the appearance of the small object causes a strong change of the local texture features, so that the local entropy value thereof also changes greatly. The background can be suppressed and the target can be enhanced by utilizing the characteristic that the appearance of the target can cause large change of the entropy value of the local image.
For each pixel point (x, y) in the infrared image I, a neighborhood space theta is set, and the size of the neighborhood space theta is mxn. Calculating the local image entropy at pixel point (x, y):
wherein is a predetermined normal number, e.g. ═ 10-6And I (I, j) represents the gray value at point (I, j) within the neighborhood Θ.
And traversing each pixel point in the infrared image I to obtain the local image entropy E of the infrared image I (as shown in C of FIG. 2). There is a homogeneous region in a of fig. 2, which has a larger entropy value according to the maximum entropy principle, such as the white region shown in C of fig. 2, but the presence of the object causes a change in the gray feature of a local region of the image, which is still visible in C of fig. 2.
Step 3, solving the weighted local image entropy H of the image:
the multi-scale gray scale difference D (shown as B in fig. 2) and the local image entropy E (shown as C in fig. 2) of the infrared image I can both realize background suppression and target enhancement on the infrared image. And D and E are fused, so that the background of the infrared image is further suppressed, and the target is further enhanced.
Performing dot product operation on the multi-scale gray difference D obtained by the processing of the step 1 and the local image entropy E obtained by the processing of the step 2 corresponding to each pixel point (x, y) to obtain a weighted local image entropy H corresponding to each pixel point (x, y), and further inhibiting the background of the infrared image and further enhancing the target, namely
The weighted local image entropy H of the infrared image I is shown in D of fig. 2. As can be seen from D of fig. 2, the background of the infrared image I is well suppressed and the target is also well enhanced.
Step 4, solving an adaptive threshold value T:
and (4) solving an adaptive threshold T for the weighted local image entropy H obtained through the processing of the steps 1,2 and 3, and carrying out binarization on the weighted local image entropy H through the adaptive threshold T to detect the infrared small target (the binarization result is shown as E in fig. 2). The adaptive threshold value T is determined by
T=c·SNR·σ+mm,SNR=(Hmax-mm)/σ (5)
Where c is a positive constant, σ is the standard deviation of the weighted local image entropy H, mm is the mean of the weighted local image entropy H, HmaxIs the maximum value of the weighted local image entropy H.
The processing results of different Infrared image processing methods are shown in FIG. 3, and it can be seen from FIG. 3 that the method of the present embodiment achieves the best results, wherein the maximum background prediction model method is from the literature (H.Deng and J.G.Liu, Infrared small target detection based on the selection-information map, Infrared Physics & Technology,2011,54(2): 100. quadrature. 107.), and the top-cap operator method is from the literature (X.Z.Bai and F.G.ZHou, Analysis of new top-hat transformation and application of free small target detection, Pattern Recognition 2010,43(6): 2145. 2156.).
The filtering effect of different infrared image processing methods (expression of SNR refers to equation (5)) is objectively evaluated using a signal-to-noise ratio (SNR)). Specific values are shown in table 1.
Table 1 SNR comparison of filtering effects using different infrared image processing methods.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.