CN106384344A - Sea-surface ship object detecting and extracting method of optical remote sensing image - Google Patents

Abstract

The invention discloses a sea-surface ship object detecting and extracting method of an optical remote sensing image, and aims at reducing the false alarm rate effectively, extracting ship objects of different sizes rapidly and accurately, obtaining amount and position information of the objects, and being low in computing complexity. Multi-vision significance is detected on the basis of a frequency-domain model, a hyper complex frequency domain transformation model and a quaternion Fourier transform phase spectral module are fused in a weighted manner to overcome disadvantages of the two models and enhance advantages of the two models, and further sea-surface background interface is inhibited, the integral continuity of detected objects and differentiation performance among the objects are enhanced, and the target area of the sea surface is searched effectively. False alarm against possible heavy cloud layers and islands in the images is reduced, an improved histogram in the gradient direction is used to represent the distribution feature of the gradient structure of the object, the detected objects are discriminated according to established rules and conditions, whether a detected object is a ship is determined, the false alarm rate is reduced greatly, and the detecting accuracy is improved.

Description

A kind of remote sensing image surface vessel target detection and extracting method

Technical field

The present invention relates to remote sensing image processing and analysis technical field are and in particular to a kind of remote sensing image surface vessel Target detection and extracting method.

Background technology

In recent years, developing rapidly with earth observation technology, the optical remote sensing imaging satellite of large quantities of high spatial resolutions Emerge in large numbers, such as：SPOT-5, IKONOS, GeoEye, Quickbird, WorldView, Pleiades series, Skysat series etc., can Obtain satellite sub-meter grade resolution panchromatic image；And the aviation image such as unmanned plane more achievable near-earth high accuracy Target Acquisition, space flight Air remote sensing provides extremely abundant data source for marine site Target detection and identification.And naval vessel is beaten as marine monitoring and wartime The important goal hit, is detected and is identified the naval vessel distribution situation that can supervise important sea areas to it, grasps enemy and fights in fact Power and analysis naval warfare information, carry out precise guidance etc.；Furthermore it is also possible to meeting maritime traffic supervision, anti-smuggling, marine searching Rescue, extend the needs of the real works such as sea fisheries, preventing and treating marine pollution.

However, in actual remote sensing image, due to lacking the priori of target and background, remote sensing shooting distance Far, and with respect to Ship Target, Sea background is complicated, particularly under severe weather conditions, cloud layer, sea fog masking, and uneven illumination Even grade affects seriously, to lead to image quality decrease, surface vessel target is faint；And sea clutter, sea surface reflection light, shade, naval vessel Tail and other interference factors (as rubbish floating object, small-sized island etc.) also can affect testing result it is easy to lead to false-alarm and Missing inspection；In addition, ship type and marine material are various, the black and white polarity of Ship Target, the different parts of same Ship Target are bright How quickly, smart degree and textural characteristics have been likely to very big difference, and these bring challenge all to surface vessel target detection, Ship Target region in standard, robustly detection and extraction marine background, wins more likely many reactions and process time, becomes For a difficult problem currently in the urgent need to address.

The extracting method being presently used for surface vessel target area is mainly summarized as follows：Traditional optical remote sensing images naval vessel Detection is mostly based on the dividing method of gray-scale statistical characteristics and marginal information, and this kind of method is applied to sea calmness, and texture is equal Even, water body gray scale is relatively low and naval vessel with sea contrast situation when larger, but to complex situations for example big wave, cloud cover, Plus naval vessel black and white polarity etc. it is easy to lead to false-alarm；Based on the method for fractal model and fuzzy theory, available natural background Detected with the difference of artificial background fractal characteristic, but when cloud and mist disturbs, background self-similarity is reduced, fractal model matching Error is larger；Method based on machine learning although various complex environment backgrounds can be tackled, but such method often relate to many Plant extraction and the search coupling of complex characteristic, time-consuming and difficulty is larger to be difficult to meet the quick requirement processing；View-based access control model notes The object detection method of mechanism, can be aware of rapidly the information related to current scene and task, be broadly divided into spatial domain and frequency Domain model, Space category model method is to extract the various features of image and merge and carry out target detection, but is applied to optical remote sensing During Ship Target Detection, target is relatively small, disturbed by sea situation, weather, illumination etc., and background rejection ability is weaker, Time-consuming relatively large, and the significance detection method based on frequency domain has a clear superiority in terms of calculating speed and background suppression, But ga s safety degree can not be simultaneous between the target when Regular value to target (when especially Ship Target is larger) and target are close Turn round and look at.In addition, when there is the interference such as thick and heavy cloud layer and island, the false alarm rate of preceding method can greatly improve although having Method for going cloud interference, but still thick and heavy cloud layer can not be removed.

Content of the invention

In view of this, the invention provides a kind of remote sensing image surface vessel target detection and extracting method, by thick Achieve the detection of surface vessel target and identification in remote sensing image to smart, can effectively reduce false alarm rate, quickly accurate Really extract different size of Ship Target, obtain its quantity and positional information, and computation complexity is low.

The remote sensing image surface vessel target detection of the present invention and extracting method, comprise the steps：

Step 1, is respectively adopted based on the supercomplex frequency-domain transform method in selective visual attention Mechanism Model and quaternary Number Fourier transformation phase place spectrometry detects to the sea-surface target of optical remote sensing original image, obtains Saliency maps S respectively₁ (x, y) and S₂(x,y)；

Step 2, to S₁(x, y) and S₂(x, y) is weighted merging, and obtains notable figure S (x, y)；

Step 3, carries out binary segmentation to notable figure S (x, y), obtains bianry image, obtains target area and Sea background Region；

Step 4, extracts target slice, specifically includes following sub-step：

Step 4.1：For each target area in bianry image, calculate minimum enclosed rectangle, wherein, boundary rectangle Height and width are parallel with the Y-axis of bianry image and X-axis respectively, then navigate to boundary rectangle in optical remote sensing original image, obtain Obtain suspected target region；

Step 4.2, described suspected target region is outwards expanded after N number of pixel, extracts from optical remote sensing original image Out, as target slice, wherein, 8≤N≤12；

Step 5, for each target slice, carries out zonule GrabCut fine segmentation；

Step 6, carries out 0 °～180 ° of Radon conversion to the target slice after fine segmentation, and maximum Radon value is corresponding Angle is the angle theta ' in cutting into slices between the major axes orientation of target and Y-axis, and the target in section is rotated described θ ' degree, obtains master Axle is parallel with Y-axis and with regard to main axisymmetric suspected target；

Step 7, characterizes target characteristic using Gradient distribution histogram method, carries out the discriminating of Ship Target and pseudo- target, tool Body includes following sub-step：

Step 7.1：The suspected target obtaining for step 6, solves the gradient in its 360 ° of directions, and by 360 ° of gradient side To being averagely divided into 8 intervals, each interval range is 45 °, be followed successively by [- 22.5 °～22.5 °), [22.5 °～ 47.5 °) ..., [- 292.5 °～337.5 °)；

Step 7.2, the suspected target that step 6 is obtained is divided into upper and lower two parts, calculates target entirety B1, target respectively The latter half B2 and target top half B3 tri- partial target image gradient amplitude statistical nature on 8 Direction intervals；

Step 7.3：Judge gradient amplitude Nogata on 8 gradient direction intervals for B1, B2 and B3 tri- partial target image Whether figure meets following condition simultaneously：

1) first and the 5th the statistical value in interval be higher than other interval；

2) first and the 5th interval symmetrical, approximately contour；

If met then it is assumed that the target of this target slice is naval vessel simultaneously, otherwise, it is not naval vessel.

Further, in described step 1, supercomplex frequency-domain transform method is improved, using the supercomplex after improving Frequency-domain transform method obtains Saliency maps S₁(x, y), specifically includes following sub-step：

Step 1.1.1, remote sensing image is transformed in CIE Lab color space, and solves CIE Lab color space In three feature passages average, and each pixel difference with three feature passage averages respectively, with described difference square As three color characteristics of remote sensing image, construct the quaternary number of each pixel position in remote sensing image；

Step 1.1.2, the quaternary number that step 1.1.1 is obtained makees discrete cosine transform, obtains the frequency domain value Q of image₁[u, v]；

Step 1.1.3, carries out smothing filtering with different gaussian kernel functions to the amplitude spectrum after discrete cosine transform, suppression High-frequency information, strengthens low-frequency information；

Step 1.1.4, the filter result that step 1.1.3 is obtained carries out inverse discrete cosine transform, obtains k Saliency maps S₁, S₂..., S_k, best scale notable figure S ' (x, y) is selected according to entropy minimum criteria, S ' (x, y) is smoothed, obtain notable Figure S₁(x,y).

Further, in described step 1, quaternary number Fourier transformation phase place spectrometry is improved, after improving Quaternary number Fourier transformation phase place spectrometry obtains Saliency maps S₂(x, y), specifically includes following sub-step：

Step 1.2.1, using three color characteristics in Lab color space, constructs each pixel in remote sensing image Quaternary number at position；

Step 1.2.2, the quaternary number that step 1.2.1 is obtained makees discrete cosine transform, obtains the frequency domain value Q of image₂[u, v]；

Step 1.2.3, calculates the logarithm value of the amplitude spectrum after discrete cosine transform；

Step 1.2.4, carries out inverse discrete cosine transform to the frequency domain value with phase information and amplitude information, after smoothing To notable figure S₂(x,y).

Further, in described step 2, first by S before fusion₁(x, y) and S₂(x, y) is normalized respectively, obtains S₁′ (x, y) and S₂' (x, y), is merged using following formula：

S (x, y)=w₁·S₁′(x,y)+w₂·S₂′(x,y)

Wherein, w₁=0.3, w₂=1-w₁.

Further, in described step 7.3, make H={ h_i, i=1,2,3 ..., 8 }, H_f={ h₁,h₅, H_p={ h₂,h₃, h₄,h₆,h₇,h₈, wherein, h_iFor 8 direction gradient statistical values in rectangular histogram,It is H_fMeansigma methodss,It is H_pAverage Value；If the gradient magnitude statistic of B1, B2 and B3 tri- partial target meets following condition simultaneously：

$\left(1\right)---\frac{\stackrel{\&OverBar;}{H_p}}{min\left(H_f\right)}<{\mathrm{\&alpha;}}_1;\left(2\right)---\frac{max\left(H_p\right)}{\max \left(H_f\right)}<{\mathrm{\&alpha;}}_2;\left(3\right)---\frac{min\left(H_f\right)}{\max \left(H_f\right)}>\mathrm{\&gamma;};$

Then can determine that the target detecting is naval vessel, be not otherwise naval vessel；Wherein, α₁、α₂It is to loosen the factor, α with γ₁= 0.6, α₂=0.7, γ=0.65.

Beneficial effect：

The inventive method does not have the parameter setting of many complexity, is independent of Sea background and the priori of target distribution characteristic yet Knowledge, the feature for Ship Target under Sea background it is proposed that the many vision significances of the combination based on frequency-domain model detection Method, improved supercomplex frequency domain transform model is merged with improved quaternary number Fourier transformation phase place spectrometry model-weight and repaiies Just respective deficiency, strengthens the advantage of the two, thus suppressing Sea background to disturb, enhances the target entirety seriality detecting Ga s safety degree and between target, efficiently searches for sea-surface target region.To the thick and heavy cloud layer being likely to occur in image and island etc. False-alarm, using improved gradient orientation histogram method characterize target gradient-structure distribution characteristicss, according to formulate rule and Condition differentiates to the target detecting, and judges whether the target detecting is naval vessel, greatly reduces false alarm rate, improves Detection accuracy.In addition, the inventive method detection and discriminating time are all second level, real-time is good, in correction rate and automatization Have in degree and be obviously improved, the Ship Target enabling marine site on a large scale quickly finds that positioning and quantity determine, the method is Calculate the informations such as the position on each naval vessel, course further combined with unmanned plane or satellite platform attitude data, and naval vessel The Classification and Identification of target lays the first stone.

Brief description

Fig. 1 is the inventive method schematic flow sheet.

Fig. 2 is surface vessel target area of the present invention pre-detection procedure chart.Wherein, Fig. 2 (a) is to input various Sea backgrounds Under optical remote sensing image；Fig. 2 (b) is that the Weighted Fusion being obtained using many vision significances model proposed by the present invention is notable Figure；Fig. 2 (c) is the result of labelling in original remote sensing images to notable figure adaptivenon-uniform sampling binaryzation；Fig. 2 (d) is target area Domain minimum enclosed rectangle calculates；Fig. 2 (e) is the target area section extracting.

Fig. 3 is Radon conversion process explanatory diagram.Wherein, the Ship Target section after Fig. 3 (a) is fine segmentation；Fig. 3 (b) For binaryzation target slice；Fig. 3 (c) converts in 0 °～180 ° of Radon for target slice；When Fig. 3 (d) is for Radon value maximum The distribution histogram being expert at corresponding projected angle；Correspond between target major axes orientation and Y-axis when Fig. 3 (e) is for Radon value maximum Angle theta '；Fig. 3 (f) is this angle that turns clockwise, and obtains symmetrical gray scale target.

Fig. 4 is improved gradient distribution schematic diagram.Wherein, in Fig. 4 (a), gradient direction (0 °～360 °) is average It is divided into 8 Direction intervals, is represented with 1D-8D, each Direction interval represents 45 °；In Fig. 4 (b), treated target is cut Piece presses B1, and region division shown in B2, B3 is three parts, and B1 is that overall goals are asked with gradient distribution, and B2 is under target Half part seeks gradient distribution, and B3 is that the top half to target seeks gradient distribution.

Specific embodiment

Develop simultaneously embodiment below in conjunction with the accompanying drawings, describes the present invention.

The invention provides a kind of remote sensing image surface vessel target detection and extracting method, introduce vision significance With gradient distribution, realize unsupervised surface vessel target area respectively and extract and goal verification, the letter of the method parameter setting Single, computation complexity is low, can effectively reduce false alarm rate, rapidly and accurately extract different size of Ship Target, obtain its quantity And positional information.

It is WINDOWS 2007 that present embodiment adopts operating system, and processor is Intel i3-2120, and dominant frequency is 3.30GHz, inside saves as 4.00GB, and experiment software processes platform is Matlab 2010a, VS 2010.Fig. 1 is optics of the present invention Remote sensing surface vessel target detection and the process flow block diagram of extracting method, Fig. 2 is sea suspected target region detection procedure chart. In conjunction with the part detection sample figure explanation in the structural framing and Fig. 2 of Fig. 1, the present invention comprises the following steps that：

Step 1：Input space resolution is optical remote sensing image f (x, y) of N × M, has naval vessel in image, and sea fog is thick and heavy Cloud layer, the size on island etc., wherein Ship Target and color polarity are different, and position distribution across the sea is also very random, As shown in Fig. 2 (a).

Step 2：Target area inspection is carried out to optical remote sensing image original image based on selective visual attention Mechanism Model Survey, search for surface vessel target area, interregional in order to strengthen the target area Regular value detecting and different target Ga s safety degree, in conjunction with multiple vision noticing mechanism models, detects to remote sensing image sea-surface target, the doubtful warship of preextraction Ship target area.

This step detailed process and computing formula are as follows：

Step 2.1：Using distant to optics based on the supercomplex frequency-domain transform method in selective visual attention Mechanism Model The sea-surface target of sense original image is detected, obtains Saliency maps S respectively₁(x,y).

Further, the present invention further improves supercomplex frequency domain from color characteristic, frequency domain transform and scale selection Alternative approach (Hypercomplex Frequency Domain Transform, HFT).

Step 2.1.1：The Lab color mode that more meet human eye subjective sensation is used as feature, by the figure in former method As rgb color space conversion to CIE Lab color space, solve the L of image, each feature passage of a, b is in entire image Average L_m, a_m, b_m：

$\left.\begin{array}{c}{L}_m=\frac{1}{N\&times;M}{\&Sigma;}_{y=1}^M{\&Sigma;}_{x=1}^{N}L(x,y)\\ a_m=\frac{1}{N\&times;M}{\&Sigma;}_{y=1}^M{\&Sigma;}_{x=1}^{N}a(x,y)\\ b_m=\frac{1}{N\&times;M}{\&Sigma;}_{y=1}^M{\&Sigma;}_{x=1}^{N}b(x,y)\end{array}\right\}---\left(1\right)$

Calculate each pixel distance with entire image each feature passage average under each feature passage in image, by original image Relevant position pixel L (x, y), a (x, y), it is squared that b (x, y) value deducts meansigma methodss successively, obtains：

$\left.\begin{array}{c}{L}_d(x,y)={\left(L\right(x,y)-L_m)}^2\\ a_d(x,y)={\left(a\right(x,y)-a_m)}^2\\ b_d(x,y)={\left(b\right(x,y)-b_m)}^2\end{array}\right\}---\left(2\right)$

By L_d, a_d, b_dAs three color characteristics of input picture, then in input picture each pixel position quaternary Number is represented by：

Q (x, y)=L_d(x,y)u₁+a_d(x,y)u₂+b_d(x,y)u₃(3)

In formula, u₁, u₂, u₃It is the base of quaternary number, meet condition：u₁ ²=u₂ ²=u₃ ²=-1, u₁⊥u₂, u₂⊥u₃, u₃⊥u₁, u₃=u₁u₂, quaternary number characteristic sequence do not affect significance calculate.

Step 2.1.2：Replace Fourier transformation using discrete cosine transform, discrete cosine transform () is made to quaternary number：

Q [u, v]=DCT (q (x, y)) (4)

Step 2.1.3：With different gaussian kernel functions, amplitude spectrum is carried out with smothing filtering, suppresses high-frequency information, strengthen low Frequency information, constitutes a spectrum metric space, the amplitude spectrum A after frequency domain transform, gaussian kernel after different amplitude spectrum smothing filterings Function G and spectrum metric space Λ is defined as follows formula：

A (u, v)=| Q [u, v] | (5)

$G(u,v;k)=\frac{1}{\sqrt{2\mathrm{\&pi;}}{2}^{k-1}{t}_0}{e}^{-(u^2+v^2)/\left(2^{2k-3}\right)}---\left(6\right)$

Λ(u,v；K)=(G (..,；k)A)(u,v) (7)

In formula：K is space scale parameter it is contemplated that the effect of the real-time that processes and actual experiment, and k can take 3～5, Calculate three yardsticks in the present invention, make k=1,2,3.Carry out inverse discrete cosine transform again and can obtain k Saliency maps S₁, S₂..., S_k, best scale notable figure S ' (x, y) is selected according to entropy minimum criteria, smooths notable figure S after being improved₁(x,y)：

S₁(x, y)=g* (S ' (x, y))²(8)

In formula, g is a multiple dimensioned gaussian kernel function.

Step 2.2：For projecting Ship Target marginal information, increase the ga s safety degree between target, based on selective visual note Meaning Mechanism Model in quaternary number Fourier transformation phase place spectrometry (Phase Quanternion Fourier Transform, PQFT) sea-surface target of optical remote sensing original image is detected, obtain Saliency maps S₂(x,y).

Further, the present invention from color characteristic, frequency domain transform and increase frequency domain amplitude information three in terms of to quaternary number Fourier transformation phase place spectrometry has done further improvement.

Step 2.2.1：In innovatory algorithm, the broad sense tuned color pattern in former method is changed to and human visual perception More close Lab color mode, directly using L, tri- color characteristics of a, b construct quaternary number：

Q (x, y)=L (x, y) u₁+a(x,y)u₂+b(x,y)u₃(9)

Step 2.2.2：Discrete Fourier transform is changed to discrete cosine transform：

Q [u, v]=DCT (q (x, y)) (10)

Step 2.2.3：Only consider in former PQFT method that phase information is different it is believed that range value is " 1 ", and present invention consideration Amplitude information AL, calculates the logarithm value of the amplitude spectrum after discrete cosine transform：

AL=lg (A)=lg (| Q [u, v] |) (11)

Inverse discrete cosine transformation, spatial domain obtains notable figure S after smoothing₂(x,y).

Step 2.3：By S₂(x, y) and S₁(x, y) Weighted Fusion, obtains notable figure S (x, y).

Wherein it is possible to select following manner to be merged：First two parts notable figure is normalized respectively before fusion, obtains To S₁' (x, y) and S₂' (x, y), is then merged using following fusion formula：

S (x, y)=w₁·S₁′(x,y)+w₂·S₂′(x,y) (12)

The notable figure calculating in view of improved HFT algorithm can be in larger plaque-like, so after many experiments, if w₁= 0.3, w₂=1-w₁, both can meet effective district partial objectives for, target can be made again to keep complete detection demand, after significance detection Shown in notable figure such as Fig. 2 (b).

Step 3：In addition it is also necessary to target is distinguished, according to notable figure by a rational threshold value with background after obtaining notable figure Gray-scale statistical property calculation threshold value T, binary segmentation is carried out to notable figure.

$S(x,y)=\left\{\begin{array}{cc}1,& S(x,y)\&GreaterEqual;T\\ 0,& S(x,y)<T\end{array}\right.---\left(13\right)$

Calculate adaptivenon-uniform sampling threshold value T using OTSU method (Da-Jin algorithm) in the present invention to realize, to notable figure coarse segmentation, being worth Region representation target area equal to 1, the region representation Sea background region that value is equal to 0, target and sea that separation detection arrives Background, after segmentation binaryzation, obtains bianry image, such as shown in Fig. 2 (c).

Step 4：Extract target slice.

This step detailed process is as follows：

Step 4.1：To in bianry image each separate target area calculate minimum enclosed rectangle, the height of boundary rectangle and Wide parallel with the Y-axis of bianry image and X-axis respectively, navigate in original remote sensing image, is marked in suspected target region it is indicated that The destination number detecting, such as shown in Fig. 2 (d).

Step 4.2：Calculate the barycenter of boundary rectangle and long width values, can be by every piece of target area from former remote sensing image Extract, when extracting, the transverse and longitudinal coordinate of each target slice is expanded on original coordinates position the individual picture of N (8≤N≤12) Element, to ensure that the target detecting has good integrity, and conveniently automatically extracts target area in follow-up fine segmentation The prospect scope in domain, through this step, can get shown in the target slice such as Fig. 2 (e) in altimetric image to be checked.

Step 5：Zonule GrabCut fine segmentation is carried out one by one to the section of multiple single goal regions.

Defining abscissa in target slice is (n, col-n), and vertical coordinate is the region in the range of (n, row-n) is prospect, Remaining region is the background of section, wherein：Col is the width of image, and row is the height of image, n=" (N/2-1), " " " for taking downwards Integral symbol.The coordinate being obtained by the method can auto-initiation prospect rectangle frame position because the target that the present invention obtains Section only comprises single target and surrounding fraction subzone, generally can be set to once by segmentation iterationses, such as Fruit sea environment is complicated, or target slice larger (more than 40 × 40), then set iterative parameter as twice, thus realizing non-friendship Mutually formula GrabCut segmentation.

Step 6：Target slice after fine segmentation is carried out with 0 °～180 ° of Radon conversion, maximum Radon value is corresponding Angle is the angle theta ' in cutting into slices between the major axes orientation of target and Y-axis, changes the major axes orientation of target in section, obtains main shaft Parallel with Y-axis and with regard to main axisymmetric suspected target.

Detailed process is as shown in Figure 3：A () is fine segmentation and the target slice of gray processing；B () cuts for binaryzation target Piece；C () converts in 0 °～180 ° of Radon for target slice, in figure most bright value corresponding abscissa angle is target projection Angle when maximum, namely target slice major axes orientation；D distribution that () is expert at corresponding projected angle when maximum by Radon value Rectangular histogram；Angle theta ' between target major axes orientation and Y-axis is correspond to when () is for Radon value maximum e；F () is this angle that turns clockwise Degree, obtains symmetrical gray scale target.

Through this step, can get a series of regularization target slice having and being vertically oriented to.

Step 7：Because the Ship Target detecting is in strip, naval vessel both sides string is symmetrical, shape facility relatively rule, And in irregular shape after jamming target (as cloud layer, island, rubbish floating object etc.) fine segmentation, so the present invention devises one kind Novel Gradient distribution histogram method (Histogram of oriented gradient, Hog) characterizes target characteristic, to rule Then change target slice and calculate gradient, and carry out the judgement of Ship Target and pseudo- target.

Detailed process is as follows：

Step 7.1：After obtaining with regard to the symmetrical target slice of Y-axis, solve its gradient, in order to preferably project target Gradient direction characteristic, 360 ° of gradient direction is averagely divided into 8 interval bins, each interval bin scope is 45 °, according to Secondary for [- 22.5 °～22.5 °), [22.5 °～47.5 °) ..., [- 292.5 °～337.5 °), shown in such as Fig. 4 (a).

Step 7.2：For more sufficiently accurately differentiating target, exclude the pseudo- target similar to naval vessel, will be through symmetrical treatment Target be sub-divided into three parts B1, B2, B3, B1 are that target is overall, and B2 is target the latter half, and B3 is target top half. As shown in Fig. 4 (b), to three parts B1, B2, B3 target image calculates its gradient amplitude on 8 Direction interval bins respectively Statistical nature.

Step 7.3：Typically, can confirm whether the target detecting is naval vessel by rule as follows：

1) statistical value of 1bin and 5bin should be higher than other bins；

2) 1bin and 5bin should be symmetrical, approximately contour；

3) B1, B2, the B3 gradient amplitude rectangular histogram on 8 gradient direction intervals, both the above condition should be met simultaneously.

Strictly the target detecting can be confirmed using mentioned above principle, but in actual applications, optical remote sensing figure As the interference such as much noise and illumination variation may be subject to, so the histogram of gradients not necessarily Striking symmetry calculating, this is sent out Bright by mentioned above principle adapt to loosen, introduce loosen the factor：α₁, α₂, γ, make H={ h_i, i=1,2,3 ..., 8 }, H_f={ h₁,h₅, H_p ={ h₂,h₃,h₄,h₆,h₇,h₈, h_i(i=1 ..., 8) is eight direction gradient statistical values in rectangular histogram,It is H_fMeansigma methodss,It is H_pMeansigma methodss, make it meet following condition： According to the Ship Target test in great amount of images, take：α₁=0.6, α₂=0.7, γ=0.65, if target B1, B2, B3 tri- Partial gradient magnitude statistic meets above three condition simultaneously, you can judge that the target that detects, as naval vessel, is not otherwise Naval vessel.

Step 7.4：It is not the zone marker deletion of Ship Target, be that the region of Ship Target retains and re-flags, defeated Go out final detection result.

It should be noted that the step shown in the FB(flow block) of accompanying drawing can be in such as one group of computer executable instructions Computer system in execute.

Obviously, those skilled in the art should be understood that each module of the above-mentioned present invention or each step can be with general Computing device realizing, they can concentrate on single computing device, or be distributed in multiple computing devices and formed Network on, alternatively, they can be realized with the executable program code of computing device, it is thus possible to they are stored To be executed by computing device in the storage device, or they be fabricated to each integrated circuit modules respectively, or by they In multiple modules or step be fabricated to single integrated circuit module to realize.So, the present invention be not restricted to any specific Hardware and software combines.

In sum, these are only presently preferred embodiments of the present invention, be not intended to limit protection scope of the present invention. All any modification, equivalent substitution and improvement within the spirit and principles in the present invention, made etc., should be included in the present invention's Within protection domain.

Claims (5)

1. a kind of remote sensing image surface vessel target detection with extracting method it is characterised in that comprising the steps：

Step 1, is respectively adopted based on the supercomplex frequency-domain transform method in selective visual attention Mechanism Model and quaternary number Fu In leaf transformation phase place spectrometry the sea-surface target of optical remote sensing original image is detected, respectively obtain Saliency maps S₁(x,y) And S₂(x,y)；

Step 2, to S₁(x, y) and S₂(x, y) is weighted merging, and obtains notable figure S (x, y)；

Step 3, carries out binary segmentation to notable figure S (x, y), obtains bianry image, obtains target area and Sea background region；

Step 4, extracts target slice, specifically includes following sub-step：

Step 4.1：For each target area in bianry image, calculate minimum enclosed rectangle, wherein, the height of boundary rectangle and Wide parallel with the Y-axis of bianry image and X-axis respectively, then boundary rectangle is navigated in optical remote sensing original image, obtain and doubt Like target area；

Step 4.2, described suspected target region is outwards expanded after N number of pixel, extracts from optical remote sensing original image, As target slice, wherein, 8≤N≤12；

Step 5, for each target slice, carries out zonule GrabCut fine segmentation；

Step 6, carries out 0 °～180 ° of Radon conversion, the corresponding angle of maximum Radon value to the target slice after fine segmentation Be the angle theta ' between the major axes orientation of target and Y-axis in cutting into slices, the target in section rotated described θ ' degree, obtain main shaft with Y-axis is parallel and with regard to main axisymmetric suspected target；

Step 7, characterizes target characteristic using Gradient distribution histogram method, carries out the discriminating of Ship Target and pseudo- target, concrete bag Include following sub-step：

Step 7.1：The suspected target obtaining for step 6, solves the gradient in its 360 ° of directions, and 360 ° of gradient direction is put down All it is divided into 8 intervals, each interval range is 45 °, be followed successively by [- 22.5 °～22.5 °), [22.5 °～47.5 °) ..., [- 292.5 °～337.5 °)；

Step 7.2, the suspected target that step 6 is obtained is divided into upper and lower two parts, calculates target entirety B1, target lower half respectively The part B2 and target top half B3 tri- partial target image gradient amplitude statistical nature on 8 Direction intervals；

Step 7.3：Judge that gradient amplitude rectangular histogram on 8 gradient direction intervals for B1, B2 and B3 tri- partial target image is No meet following condition simultaneously：

1) first and the 5th the statistical value in interval be higher than other interval；

2) first and the 5th interval symmetrical, approximately contour；

If met then it is assumed that the target of this target slice is naval vessel simultaneously, otherwise, it is not naval vessel.

2. remote sensing image surface vessel target detection as claimed in claim 1 with extracting method it is characterised in that described In step 1, supercomplex frequency-domain transform method is improved, significance is obtained using the supercomplex frequency-domain transform method after improving Figure S₁(x, y), specifically includes following sub-step：

Step 1.1.1, remote sensing image is transformed in CIE Lab color space, and solves three in CIE Lab color space The average of individual feature passage, and each pixel difference with three feature passage averages respectively, using described difference square as Three color characteristics of remote sensing image, the quaternary number of each pixel position in construction remote sensing image；

Step 1.1.2, the quaternary number that step 1.1.1 is obtained makees discrete cosine transform, obtains the frequency domain value Q of image₁[u,v]；

Step 1.1.3, carries out smothing filtering with different gaussian kernel functions to the amplitude spectrum after discrete cosine transform, suppresses high frequency Information, strengthens low-frequency information；

Step 1.1.4, the filter result that step 1.1.3 is obtained carries out inverse discrete cosine transform, obtains k Saliency maps S₁, S₂..., S_k, best scale notable figure S ' (x, y) is selected according to entropy minimum criteria, S ' (x, y) is smoothed, obtains notable figure S₁(x,y).

3. remote sensing image surface vessel target detection as claimed in claim 1 with extracting method it is characterised in that described In step 1, quaternary number Fourier transformation phase place spectrometry is improved, using the quaternary number Fourier transformation phase spectrum after improving Method obtains Saliency maps S₂(x, y), specifically includes following sub-step：

Step 1.2.1, using three color characteristics in Lab color space, constructs each location of pixels in remote sensing image The quaternary number at place；

Step 1.2.2, the quaternary number that step 1.2.1 is obtained makees discrete cosine transform, obtains the frequency domain value Q of image₂[u,v]；

Step 1.2.3, calculates the logarithm value of the amplitude spectrum after discrete cosine transform；

Step 1.2.4, carries out inverse discrete cosine transform to the frequency domain value with phase information and amplitude information, is shown after smoothing Write figure S₂(x,y).

4. remote sensing image surface vessel target detection as claimed in claim 1 with extracting method it is characterised in that described In step 2, first by S before fusion₁(x, y) and S₂(x, y) is normalized respectively, obtains S₁' (x, y) and S₂' (x, y), under adopting Formula is merged：

S (x, y)=w₁·S₁′(x,y)+w₂·S₂′(x,y)

Wherein, w₁=0.3, w₂=1-w₁.

5. remote sensing image surface vessel target detection as claimed in claim 1 with extracting method it is characterised in that described In step 7.3, make H={ h_i, i=1,2,3 ..., 8 }, H_f={ h₁,h₅, H_p={ h₂,h₃,h₄,h₆,h₇,h₈, wherein, h_iFor straight Square 8 direction gradient statistical values of in figure,It is H_fMeansigma methodss,It is H_pMeansigma methodss；If B1, B2 and B3 tri- part mesh Target gradient magnitude statistic meets following condition simultaneously：

$\left(1\right)---\frac{\stackrel{\&OverBar;}{H_p}}{\min \left(H_f\right)}<{\mathrm{\&alpha;}}_1;\left(2\right)---\frac{\max \left(H_p\right)}{\max \left(H_f\right)}<{\mathrm{\&alpha;}}_2;\left(3\right)---\frac{\min \left(H_f\right)}{\max \left(H_f\right)}>\mathrm{\&gamma;};$

Then can determine that the target detecting is naval vessel, be not otherwise naval vessel；Wherein, α₁、α₂It is to loosen the factor, α with γ₁=0.6, α₂ =0.7, γ=0.65.