A topic-oriented image panoramic map feature description and extraction method

Abstract

The application discloses a kind of theme-oriented image panorama pattern feature description and extraction method, it is related to computer vision technical field, including with meet complete and mutually exclusive as criterion, based on ontology theory definition panorama pattern concept understanding general model;According to gestalt theory, the general model of concept understanding of panorama pattern is described, and through panorama sample marking, sample unit active calculation and gestalt construction, constructs panorama pattern sample library, symbiotic relationship dependent factor library, gestalt pattern library;According to the one-to-two of tower piece and the one-to-many touch control strategy of inside, the construction factor in gestalt pattern library is mapped to the modeling process of pixel symbiotic feature as supervision information, and constructs pixel level pattern feature coupling extraction model.Therefore, by using the above-mentioned theme-oriented image panorama pattern feature description and extraction method, it can really cross the semantic gap, realize the full true extraction of thematic pattern feature, and conform to the logic information transmission whole process of human cognitive psychology perception.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and in particular to a thematic image panoramic pattern feature description and extraction method. BACKGROUND

[0002] Pattern is a kind of high-level perception of human visual system to the repeated spatial structure of image thematic target, which exists widely and is reflected in remote sensing images as a high-level attribute of geographic space, naturally has the comprehensive representation attribute of spatial texture and structure relationship between entities, and is an important means to describe and distinguish different thematic ground objects.

[0003] In recent years, with the continuous expansion and integration of high-resolution remote sensing image data and thematic application scenarios in time scale, spatial scale and content composition, new requirements and challenges have been put forward for pattern analysis methods: on the one hand, pattern features show thematic differentiation and ubiquitous batch characteristics, such as in land classification application, the objects of pattern description cover various innovative combinations of land use / cover thematic targets such as "from street to administrative region", "from history to present situation", "from 1st level to 3rd level"; on the other hand, the spatial structures between thematic targets show complex and overlapping characteristics, so when modeling the pattern, it is required not only to have the ability to quickly understand the diversified semantic connotation and respond to the huge task volume, but also to solve more serious problems of uncertain spatial structure such as shadow occlusion, same object different spectrum and same spectrum different object. Therefore, detailed and accurate pattern features are increasingly important in thematic analysis, and it is urgent to design a high-level description model with strong connectivity, indication, anti-interference and self-adaptation.

[0004] At present, in computer interpretation research, pattern is regarded as a kind of context feature based on image gray distribution, and data-driven methods are used to extract and understand its underlying information, including four implementation approaches: 1) statistical method based on pixel gray value, which uses LBP, HOG, SIFT, GLCM, etc. in local range to make correlation statistics on pixel gray value to obtain its spatial distribution information; 2) statistical method based on signal, which makes statistical analysis after signal conversion of original image, and uses Tamura, autoregression, wavelet transform or Gibbs algorithm to find visual mutation information in other feature spaces except original gray value; 3) abstract feature extraction method relying on deep learning, the main idea of which is to construct convolution kernels of different sizes to realize indirect description of context features through deep network calculation; 4) structure statistical method based on segmented objects, which usually uses OTSU, entropy method, etc. to divide image segmentation blocks through threshold combination with gray clustering, seed growing algorithm, and estimates fixed or semi-fixed segmentation block structure template through text statistical idea, and the commonly used models include: Voronio checkerboard, probability theme, semantic bag of words, Markov random field, etc.

[0005] However, the existing pattern feature extraction method is mainly based on visual attention theory, adopts a bottom-up calculation scheme, and emphasizes the mining of information of image data itself; on this basis, such research has to rely on a post-processing optimization step, and the bottom layer "visual semantics" is strongly processed and re-understood to approximate the corresponding relationship between visual semantics and thematic semantics information, resulting in limited representation ability of pattern features to real semantics, underestimated effectiveness, and inability to quickly and effectively solve the problem of current widespread thematic scene image cognitive analysis and batch accurate processing.

[0006] Therefore, there is an urgent need for an extraction paradigm of "panoramic combination system-Gestalt model construction-pixel mapping fusion calculation" to change the bottom logic of pattern feature extraction from data dependence to thematic mode whole process supervision, and more accurately fit the psychological perception and visual interaction process of human thematic image cognition. SUMMARY

[0007] The purpose of the present application is to provide a thematic image panoramic pattern feature description and extraction method, to explore the internal spatial correlation mechanism of pattern semantics from "panoramic domain", "tower domain" and "spectrum domain", form a "top-down and multi-core propagation" pattern feature cascade cognition and calculation idea of "landscape tower spectrum", and build a whole sample mode description and coupled extraction model based on substantive Gestalt construction and tower piece chain mapping calculation.

[0008] To achieve the above purpose, the present application provides a thematic image panoramic pattern feature description and extraction method, comprising the following steps:

[0009] S1, based on the criterion of satisfying completeness and mutual exclusion, defining a panoramic pattern concept understanding general model GO based on ontology theory, and standardizing the expression as GO={SC, SP}, SC represents a thematic full combination category system, and SP represents a panoramic pattern concept based on symbiotic relationship;

[0010] S2, according to the Gestalt theory, describing the panoramic pattern concept understanding general model, constructing a panoramic pattern sample library, a symbiotic relationship dependent factor library and a Gestalt pattern library through panoramic sample labeling, sample unit active calculation and Gestalt construction;

[0011] In the symbiotic relationship dependent factor library, the intra-class symbiotic operator includes size S and orientation C: size S represents the average value of the short side of the circumscribed rectangle of the same category, and orientation C represents the average value of the main direction of the same category; the inter-class symbiotic operator includes frequency distance D and frequency direction G: distance D represents the minimum value of the high frequency distance between different categories, and direction G represents the direction corresponding to the high frequency distance between different categories; a scene is expressed by a set of four-dimensional attribute vectors of size S, orientation C, distance D and direction G, and a scene similarity coefficient is designed using the confidence value of the four-dimensional attribute vector to measure the similarity of two scenes.

[0012] In the Gestalt pattern library, a general model is understood according to the panorama type concept, a tile is established for each scene, an equidistant grid is arranged inside the tile, and the marked points on the grid represent the symbiotic relationship, and the tile includes size, orientation, point distance, point diameter, and thickness attribute dimensions;

[0013] S3, according to the one-to-two and one-to-many touch strategies of the tile, the construction factor in the Gestalt pattern library is mapped to the pixel symbiotic feature modeling process as supervision information, and a pixel-level type feature coupling extraction model is constructed.

[0014] Further, in S1, the thematic full combination category system includes scenes, basic classes, and auxiliary classes; wherein the basic class is a substantive object defined according to the thematic task, and the auxiliary class is another substantive object that exists in dependence on the basic class; and the inter-class relationship includes parallel symbiotic relationship and parent-child symbiotic relationship;

[0015] The panorama type concept based on symbiotic relationship includes four symbiotic spatial relationships: apart symbiosis, adjacent symbiosis, containing symbiosis, and complex symbiosis.

[0016] Further, the extraction steps of the panorama type concept based on symbiotic relationship include:

[0017] According to the basic spatial relationship attribute topology Q, orientation H, metric O, and geometric shape R, four dimension attributes of the generalized region connection calculus GRCC are formed; then, the grouping superposition strategy is used for type concept combination deduction and understanding, and the topology and geometric shape relationship combination, and the orientation and metric relationship combination are extracted.

[0018] Further, in S2, the construction of the panorama type sample library includes:

[0019] According to the Gestalt whole-part theory, the scene is taken as the carrier of the type unit, and the whole sampling strategy based on the scene unit is used to mark the panorama sample.

[0020] The sources of the panorama sample include the latest global map polygons based on the GIS thematic historical database, the global map polygons based on the thematic subject interpretation, and the global map polygons integrated based on the GIS thematic historical data and remote sensing image data interpretation results.

[0021] Further, in S2, the construction of the symbiotic relationship dependent factor library includes:

[0022] First, the shape characteristics of the panorama sample map polygon are calculated: the polygon circumscribed rectangle width W and the rotation angle θ, which are used to describe the regular shape of the polygon; the fractal dimension F d , and the rectangular degree Rect, which are used to describe the confidence of the regular shape of the polygon;

[0023] Wherein, Fd = 2ln(P / k) / ln(A), P, A are the perimeter and area of the patch respectively; Rect = A / A MER , A MER is the minimum circumscribed rectangle area of the patch;

[0024] Then, according to the definition of the generalized region connection calculus, the orientation H and the metric O are expressed as the core of the pattern factor form, and the confidence probability is added to represent the relative fuzzy symbiotic relationship according to the strength of the pattern characteristics:

[0025] The intra-class symbiotic operator represents the shape similarity, including size S and direction C, which are expressed by the average value of the regular shape of the patch, as follows:

[0026] S i = averge(W);

[0027]

[0028] In the formula, subscript i represents the parameters corresponding to class i, and n represents the number of categories of the class;

[0029] The inter-class symbiotic operator is obtained by using the double histogram threshold analysis method, including the statistical metric histogram, taking the minimum value of the distance histogram peak value between different classes as the distance D, and taking the direction corresponding to the distance D as the orientation G, as follows:

[0030] D ij = min(Peak(Freq(Dis ij ))1,…,Peak(Freq(Dis ij )) k );

[0031] G ij = Dir(max(Num(D ij ) B |B = [45°, 90°, 135°, 180°]);

[0032] In the formula, D ij represents the minimum value of the distance histogram peak value between class i and class j, that is, the minimum value of the high-frequency distance; Dis ij represents the distance between the corresponding patches of class i and class j; k represents the number of distance histogram peak values between class i and class j; G ij represents the main direction corresponding to the minimum value of the minimum high-frequency distance between the classes; Freq represents the frequency histogram statistical function; Peak represents the peak value statistical function in the frequency histogram, Num represents the statistical patch number function, (D ij ) B represents D ijThe figure spot distributed in direction B, Dir represents the corresponding direction function of the statistical result.

[0033] Further, each tower piece represents a full combination parallel symbiotic relationship or a full combination parent-child symbiotic relationship; the full combination parallel symbiotic relationship is located at the tower base position, the inner network points marked inside have symmetry, and only half of the symbiotic relationship inner network points with practical significance and independence are retained; the full combination parent-child symbiotic relationship is located at the tower top position; each tower piece is staggered and stacked according to the category system and the size range to form a panoramic format tower of a number of symbiotic scene cascades.

[0034] Further, in the full combination parallel symbiotic relationship, the size of the tower piece is S=S i +max(S j +D ij ), the orientation of the tower piece is C=(C i +max(C j +G ij )) / 2, the point distance of the tower piece is D=D ij , the point diameter of the tower piece is G=G ij , and the thickness of the tower piece is V=α×β×(1+SV×CV); the symbiotic mode includes apart symbiosis and adjacent symbiosis.

[0035] Wherein,

[0036]

[0037] β=max(Num(D ij ) B ) / Num(D ij );

[0038]

[0039] In the formula, α represents the weight proportion of the number of pairs of categories i and j appearing in the statistical scale with a distance of D to the total number of high-frequency distances, Dis ij represents the distance between the figure spots of categories i and j, β represents the weight proportion of the number of pairs of categories i and j appearing in the statistical scale with a distance of D and a direction of G to the number of directions when the distance is D, SV represents the standard degree of the size S, represents the average size of all categories, and CV represents the standard degree of the orientation C.

[0040] In the full combination parent-child symbiotic relationship, the size of the tower piece is S=S i , the orientation of the tower piece is C=C i , the point distance of the tower piece is D=D ij , the point diameter of the tower piece is G=G ij , and the thickness of the tower piece is V=SV×CV×(1+α×β); the symbiotic mode includes inclusion symbiosis and complex symbiosis.

[0041] Further, the element of each scene is composed of (SV, CV, a, b), which represents the confidence probability of the corresponding attribute value appearing in the scene. The expression of the scene similarity coefficient is:

[0042]

[0043] In the formula, V(i, j) represents the scene similarity coefficient of scene Ai and scene Aj, that is, the scene confidence; F represents the sum of the products of the scene attributes.

[0044] Further, in S3, the pixel-level pattern feature coupling extraction model includes correcting the statistical results based on the scene confidence V=f(SV, CV, a, b), and improving the energy E, the contrast I, the entropy H, and the correlation J four traditional classic descriptors, as follows:

[0045] GFCM(S, C, D, G, V) i =∑ G (dG×dD) windows(S,C) ×V;

[0046] E(D, G, V) = ∑ t1 ∑ t2 [p(t1,t2|D,G)×V] 2 ;

[0047] I(D, G, V) = ∑ t1 ∑ t2 (t1-t2) 2 [p(t1,t2|D,G)×V];

[0048] H(D, G, V) = -∑ t1 ∑ t2 [p(t1,t2|D,G)×V]log(p(t1,t2|D,G)×V);

[0049]

[0050] Wherein,

[0051] p(t1,t2|D,G) = {[(x,y),(x+dx,y+dy)] ∈ (S1*S2),C|f(x,y) = t1,f(x+dx,y+dy) = t2};

[0052] μ t1 =∑ t1 t1∑ t2 p(t1,t2|D,G);

[0053] μ t2 =∑t2 t2∑ t1 p(t1,t2|D,G);

[0054] σ t1 2 =∑ t1 (t1-μ t1 ) 2 ∑ t2 p(t1,t2|D,G);

[0055] σ t2 2 =∑ t2 (t2-μ t2 ) 2 ∑ t1 p(t1,t2|D,G);

[0056] In the formula, GFCM(S,C,D,G,V)i represents the gray level co-occurrence matrix correction statistics obtained by the tile mapping on any image element, and the subscript i represents the corresponding category;(t1, t2) represents the gray level value of the point pair corresponding to the distance D and the direction G;Window(S,C) represents the observation window;(x,y) represents the position coordinates of any point in the image;(x+dx,y+dy) represents the point corresponding to the point (x,y) with the distance D and the coordinate horizontal axis angle G;dG and dD respectively represent the statistical direction and statistical distance of the gray level co-occurrence matrix;p(t1,t2|D,G) is the joint probability matrix.

[0057] Therefore, the application adopts the above-mentioned image panoramic pattern feature description and extraction method for a theme, and has the following technical effects:

[0058] (1) The calculation framework of the existing co-occurrence relationship logical semantic mode lacks completeness and consistency and has weak description ability, the application constructs the unified theory and method framework of the “ontology understanding-mode description-feature extraction” based on the format, ensures the consistency of the logical semantic expression of the pattern co-occurrence relationship, and solves the problems of incomplete high-level spatial relationship logical semantic mode cognition and the inability of direct matching with the bottom feature calculation framework.

[0059] (2) The application proposes an adaptive gestalt pattern depth mapping and dynamic transmission mechanism, which performs full-parameter heuristic and constraint modeling on pixel characteristics, accurately transmits thematic panoramic pattern samples, construction factors and typical patterns, and reduces the complexity and uncertainty of pixel pattern feature modeling. Meanwhile, the application proposes a multi-layer co-occurrence kernel modeling of pixels, which highlights the fusion of static explicit and dynamic implicit co-occurrence relationship. Compared with the traditional pyramid multi-scale modeling, the model contains a spectrum of multiple patterns, and is associated with the full-parameter modeling process of window size, orientation, statistical distance, direction, etc. The panoramic pattern cascade transmission of "one-to-two" of the tile window and "one-to-many" of the tile axis point can be realized, and the complete logical chain of pattern feature cognition is formed, which more accurately describes the advanced co-occurrence information of complex features.

[0060] (3) The application can effectively eliminate the small spot interference phenomenon of "same object different spectrum" through top pattern mapping calculation, and effectively solve the problem of "same spectrum different object" of land class boundary recognition through bottom pattern mapping calculation, realize the bottom logic of pattern feature extraction from data dependence to thematic pattern comprehensive guidance, expand the breadth and depth of real pattern feature extraction, reduce the number of bottom features while improving the accuracy and robustness, have the characteristics of structural invariance and cross-theme and cross-scene adaptability, and ensure its accurate generalization.

[0061] The technical solutions of the application will be further described in detail below with the help of the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a whole flow chart of a thematic image panoramic pattern feature description and extraction method;

[0063] Figure 2 It is a general ontology framework schematic diagram of a thematic panoramic pattern in an embodiment of a thematic image panoramic pattern feature description and extraction method;

[0064] Figure 3 It is a gestalt pattern construction schematic diagram of a thematic panoramic pattern in an embodiment of a thematic image panoramic pattern feature description and extraction method, wherein (a) is a parent-child co-occurrence gestalt tile construction schematic diagram of a class i, (b) is a panoramic pattern gestalt pattern construction schematic diagram, and (c) is a thematic panoramic pattern gestalt pattern schematic diagram;

[0065] Figure 4 It is a GFCM calculation schematic diagram of an internal street and road in an embodiment of a thematic image panoramic pattern feature description and extraction method. DETAILED DESCRIPTION

[0066] The present application will be explained in more detail by the following examples, the purpose of which is to protect all variations and improvements within the scope of the present application, and the present application is not limited to the following examples.

[0067] The problems of the existing pattern feature extraction method are as follows:

[0068] (1) The pattern feature extraction framework is bottomed and complicated. From the visual attention theory, the pattern feature is equated to the context feature, and the abstract context concept model in the general field of computer pattern recognition is simply followed. There is a fundamental logical defect that the model bottom abstraction hypothesis does not match the thematic analysis task. This conflict becomes a bottleneck problem that restricts the effectiveness of pattern feature extraction. The existing mainstream pattern model adopts an image data driven scheme from spectrum to pattern, forming an extraction paradigm of "visual semantic mining - pattern semantic understanding". Starting from the one-way inefficient cognitive strategy of sample labeling, visual attention, structure statistics, semantic assignment and task understanding, each part is calculated independently, which artificially separates the internal unity between pattern modeling process and real semantics, causing the defects of cumulative, irreversible and uncertain errors in feature description.

[0069] (2) The sample labeling of pattern is fragmented and quantified. The pattern has the property of multi-granularity complex spatial relationship, which often needs high-quality sample labeling to describe and guide feature extraction. However, the construction of samples in traditional methods depends on independent target space, emphasizes the purity of sample unit, can only describe the spatial properties of the target itself, and cannot express the specific high-level spatial relationship between targets, which has the problem of fragmentation of sample unit and its covered information. Therefore, existing research, especially deep learning algorithms, often rely on the number of tens of thousands of sample quantity to obtain more spatial attribute patterns. Although some classic sample libraries have been developed and shared by many famous research institutions, users still need to do a lot of extra work when facing different thematic tasks, which will certainly increase the burden of labeled samples and algorithm time consumption. In order to solve the problem of high labeling cost of thematic sample diversification, especially for the dimension reduction demand of existing deep learning large models, transfer learning, reinforcement learning, random forest and other auxiliary schemes are used to try to derive universal samples, models and parameters. However, in fact, it can only be verified to be effective in a small range or single category experiment. When the thematic category system and spatial relationship are updated, it is difficult to generalize and migrate, and a large amount of parameter testing and sample supplement is still needed. Moreover, these "quantity for quality" improvement strategies are extremely unstable in precision, and sometimes even cannot reach the description ability of traditional pattern features.

[0070] (3)Pattern feature extraction simplification and abstraction. In order to increase data advantages, the current mainstream pattern feature extraction algorithm mainly adopts a multi-scale strategy to fit more context features. A scale pyramid computing framework is designed to expand the pattern visual hierarchy, and the most representative one is the deep convolution operator. However, the multi-scale basic template still only relies on the experience size, regular forward and interval of 1 statistical solidification mode, without constructing an effective response model for various pattern detail differentiation characteristics, which has the problem of setting simplification and is easy to be disturbed by data noise and land cover information; and the feature expansion is at the cost of increasing the number of scales, which not only has low computational efficiency and abstract information redundancy, but also is easy to limit the local optimum or data optimum, causing scale connotation indication confusion, semantic drift, conclusion deviation and instability and other problems. A large number of researchers realize the connection between scale parameters and thematic target space, and further use feature enhancement and optimization algorithms such as pyramid pooling, up-sampling optimization, dilated convolution and expanded convolution. Unfortunately, these studies do not carry out the whole process of pattern feature modeling, and still take abstract information as the basis in actual modeling, ignoring the accurate embedding and combined application of real pattern parameters, and the number cost is still serious, while the actual semantics is still deviated.

[0071] (4)Pattern feature understanding one-sidedness and explicitness. Under the bottom-up logic of data-driven, existing models have to rely on the research idea of "post-matching" of pattern thematic information. In the "post-matching" method, traditional research mainly proposes methods such as mid-level feature bag-of-words, feature aggregation and semantic assignment based on segmentation primitives; in recent years, due to the data encapsulation and reinforcement engineering advantages of deep learning models, visual implication and domain knowledge question answering algorithms are proposed for convolution network calculation results. However, due to the differences in processing ideas, expert knowledge and text sources, they often focus on understanding some regular explicit pattern features in the theme, such as the topological and metric relationship between two classes, and use independent concept description and two-by-two combination reasoning model. Ignoring the feature concepts such as direction and shape, and their multi-granular combination reasoning and generalization use, the spatial relationship description of details is insufficient; without constructing a unified and effective concept model for various pattern features, the reasoning information lacks completeness and consistency, and the robustness and portability are poor. The construction process of various real semantic libraries is complex, but there are still problems of unclear indication, incomplete and wrong, and the bottom understanding and cognitive bias have not been substantially changed.

[0072] In view of the above situation, the present application provides a thematic image panoramic pattern feature description and extraction method for the demand of high-resolution remote sensing image ubiquitous thematic analysis and the task characteristics of complex anti-interference, including a general model of ontology concept understanding, a Gestalt description model of panoramic mode and a pixel full-parameter mapping extraction model, as shown in Figure 1 The specific steps are as follows:

[0073] S1, based on ontology theory, starting from the construction of the full combination category system based on the theme, the full combination relationship system based on the category, and the full combination sampling system based on the substantive, understand the general concept of panoramic pattern features; by establishing a pattern feature general ontology concept model and mode that meets the JEPD (Jointly Exhaustive and Pairwise Disjoint) condition, reduce the incompleteness of the understanding of the theme pattern concept.

[0074] The pattern reveals the repeated spatial structure between different ground object categories, that is, the spatial relationship concept corresponding to the analysis of the category "symbiosis". Therefore, taking the JEPD condition as the criterion, defining the panoramic pattern understanding framework GO based on ontology theory, from the perspective of standardization and ease of use, the task-driven thematic domain ontology framework is standardized as: GO = {SC, SP}, as shown in Figure 2 , quickly clarify the full combination category system SC and the full combination symbiotic relationship system SP with the theme scene as the core.

[0075] SC (scene-categories) represents the thematic full combination category system, and its concept hierarchy includes: scene, basic class, and affiliated class. Among them, the basic class is the substantive object defined according to the thematic task, and there are N of them; the affiliated class is other substantive objects that exist in dependence on the basic class, such as street trees dependent on roads, and there are M of them.

[0076] SP(scene-position) defines the panoramic pattern concept based on the co-existence relationship. Based on the description of the category system ontology, the single co-existence relationship is found on the space target whole algorithm based on the logical and rigorous RCC algebra logic, and the co-existence relationship is combined. First, the basic spatial relationship attributes are defined, including the traditional 3 attributes: topology Q, orientation H, and metric O, and the newly added geometric shape R. The co-existence relationship is expressed as direction similarity, distance similarity, and shape similarity, forming the GRCC(Generalized region connection calculus)4 attribute dimensions. On this basis, the "grouping superposition" strategy is used for pattern concept combination deduction and understanding: first, the topology and geometric shape relationship combination constructs the Q-R model, which can define the most complete and stable pattern qualitative relationship characteristics; second, the orientation and metric relationship combination constructs the H-O model, which can realize the quantitative pattern relationship characteristics definition. The two combinations complement and depend on each other, and the concept expression paths are compatible, so in this embodiment, the "Q-R model" + "H-O model" is used to define the co-existence spatial relationship of panoramic pattern, and the four combined forms of disjoint co-existence, adjacent co-existence, containing co-existence, and complex co-existence of the full combination category are extracted, and the GRCC combination table is constructed to define the real co-existence relationship, which has the advantages of concise and unified pattern concept, complete real details, and meets the JEPD conditions, as shown in Table 1.

[0077] Table 1 GRCC pattern co-existence relationship type definition

[0078]

[0079] For the Q-R combination form of pattern attributes, the ternary operation operator and the binary logic(True, False) are used for expression, then 23 kinds of geometric shape + topology relationship can be expressed. Among them, (T, T, T) and (T, T, F) two cases are impossible in the real geographical world. In addition, (F, T, F), (F, F, F) and (T, F, F) are two-by-two disjoint, intersection and containing three basic relationships, (F, T, T) represents the newly defined geometric shape spatial relationship. Based on the four basic spatial relationship attributes, the new complex co-existence relationship OSC is defined, which is actually the intersection of multiple disjoint co-existence, adjacent co-existence and containing co-existence, and the specific expression is:

[0080]

[0081] For any arbitrary map Xi, i = 1, 2, …, n, it can be inferred that Xi belongs to Yi, while any arbitrary map of X satisfies the geometric shape similarity rule, and any arbitrary map of Y also satisfies the geometric shape similarity rule; T1, T2 are the threshold values of the shape deviation of the map, and less than the corresponding threshold value, it is considered similar.

[0082] This matches the connotation of the joint satisfaction problem (JSP) of the panoramic pattern, that is, "in the geometric shape (M1, O1) range of the basic class, in the position with M2 and O2, there is always a certain "topology + shape" symbiotic phenomenon", which has the uniqueness of formal results and robust and effective supervision.

[0083] S2, based on the Gestalt theory, realizes the construction of the pattern construction factor library through panoramic sampling marking and active calculation; at the same time, based on the multi-dimensional assignment and tower superposition of the tile unit, realizes the construction of the pattern mode library, fits the real pattern information from multiple attribute dimensions of the Gestalt model, forms a comprehensive and accurate expression of thematic semantics to pattern mode, and maximizes the loss of real pattern information and the uncertainty of description.

[0084] In the Gestalt theory, the traditional pyramid model uses a regular fixed grid division to express the diversity of scales. Among them, the most basic discovery is that human vision is holistic, which automatically connects similar and close land classes in the simulation pattern vision as a group; through object grouping to perceive things, it is more inclined to perceive the form of pattern continuous isomorphism rather than fragmented information; even non-continuous things will be automatically visualized as complete. Therefore, the Gestalt pattern can undergo extensive changes without losing its own characteristics. For example, a format can still maintain the same structure after modulation, although the shape of the internal components deviates, the structure rule can still be perceived through conversion.

[0085] According to the isotype theory, the Gestalt theory, etc., determine the sample marking principle:

[0086] (1) According to the Gestalt whole-part theory, it is believed that the scene is the carrier of the pattern unit, and the whole sampling strategy based on the scene unit is adopted to simulate the behavior of taking the scene as the basic carrier of pattern understanding in visual interpretation; rely on the latest 1 period of global background data for marking, which can meet the needs of light and high-quality sample set collection, not only avoid the pattern real loss problem caused by traditional random and fragmented sampling methods, but also analyze and deliver the full combination of symbiotic relationship system.

[0087] (2) Not necessarily complete accurate labeling, allowing for sample collection error, reduce the sample time requirements and location requirements, to meet the Gestalt similarity law, proximity law as the standard, while ensuring that different experts' perceptual information, although there are quality, but is still effective.

[0088] Thus, the multi-source collection mechanism of panoramic samples is determined: first, based on the latest global plot in the GIS thematic historical database, directly construct the background data; second, based on the thematic visual interpretation of the latest image in the remote sensing historical database to obtain the global plot, indirectly construct the background data; third, based on the integration of multi-period incomplete GIS thematic historical data and remote sensing image data interpretation results, supplement, splice and update the global plot, and fuse to construct the background data.

[0089] According to the continuity law and similarity law of Gestalt, the gradual relationship mining and fuzzy expression method of confidence are used to construct the GRCC symbiotic relationship dependent factor library, as follows:

[0090] First, the shape characteristics of the panoramic sample plot are calculated, as shown in Table 2.

[0091] Table 2 Shape factor calculation based on panoramic sample plot

[0092]

[0093] Next, the pattern structure factor of the panoramic sample plot is calculated. The Gestalt similarity law indicates that the pattern mode has independence, separability, fuzziness and migration. According to the definition of GRCC, if a group of land features symbiotic arrangement, then in a certain direction H, a certain distance O, there are always class pairs. Therefore, H-O (azimuth-measure) naturally carries the semantic of spatial symbiotic relationship, and is used as the core pattern factor formal expression form; at the same time, according to the strength of the pattern characteristics, the confidence probability is added to represent the relative fuzzy symbiotic relationship.

[0094] Based on the above assumptions, this embodiment measures whether they have a symbiotic relationship through the full combination relationship of intra-class and inter-class. The main intra-class symbiotic operator is shape similarity, including size S and orientation frequency C, which is expressed by shape average value. The main inter-class symbiotic operator is distance D and direction G. Through double histogram threshold analysis method, the frequency histogram statistical function Freq is used to determine: first, the measurement histogram is calculated, and the minimum value of its peak value is taken as the frequency distance; then the peak value corresponding statistical direction histogram is drawn, and the minimum value of its peak value is taken as the frequency direction to find the inter-class symbiotic dependent factor (D, G).

[0095] For the intra-class form, S, C describe the minimum of the high-frequency factor of intra-class H-O, SV, describe the standard degree of the average S of intra-class, CV describe the standard degree of the average C factor of intra-class, these factors collectively express the size of the intra-class form of the frequency of similar S, similar C, and their confidence SV, CV; D, G describe the minimum of the high-frequency factor of inter-class H-O, α describe the minimum high-frequency factor of the inter-class H histogram, β describe the weight of the inter-class H factor, these factors collectively express the measurement of the inter-class topology of the frequency of similar D, similar G, and their confidence (α, β). In this embodiment, the scenes Ai and Aj are two sets of 4-dimensional attribute vectors (size S, orientation C, direction D, distance G), and their confidence values are (SV, CV, α, β). Therefore, the values of all set elements are not traditional 0 or 1, but fuzzy values in the value range [0, 1], and each element represents the confidence probability of the corresponding attribute value appearing in the scene.

[0096] Table 3: Panoramic sample-based pattern construction factor calculation

[0097]

[0098]

[0099] According to the GRCC pattern definition, a tile is established for each scene, the internal grid spacing of the tile is 1, and the marked points on the grid represent the symbiotic relationship. Therefore, the format pattern is constructed in multiple attribute dimensions such as the size, orientation, point distance, point diameter, and thickness of the tile, and the tiles are associated with each other. The tile unit includes tiles that form parallel symbiotic relationships in basic categories and tiles that form parent-child symbiotic relationships in subsidiary categories, as shown in Table 4.

[0100] Table 4: Format pattern construction based on panoramic sample pattern factors

[0101]

[0102] Each tile represents a full combination of parallel symbiotic relationships of a category i, located at the base of the tower, and internally marked with N internal network points; or a full combination of parent-child symbiotic relationships, located at the top of the tower, and internally marked with M internal network points. Since the parallel internal network points of the tile have symmetry, only half of the symbiotic relationship internal network points with practical significance and independence are retained: the base tile of the parallel symbiosis has N internal network points, and the symbiotic scenes expressed by the internal network points have , and the top tile of the parent-child symbiosis has N internal network points, and the symbiotic scenes expressed by the internal network points have . The above tile units are then staggered and overlapped according to the category system and the size of the range, thereby cascading multiple symbiotic scenes to form a panoramic format, as shown in Figure 3 . The total number of tiles is 2N, and the total number of scenes is One. Describes the Gestalt patterns that meet the JEPD condition and are suitable for image pattern feature calculation.

[0103] In one embodiment, the community pattern Gestalt pattern classification is obtained according to the GRCC pattern Gestalt pattern, as shown in Tables 5 and 6.

[0104] Table 5 Community thematic pattern Gestalt pattern classification defined based on panoramic combination concept

[0105]

[0106] Table 6 Community thematic pattern and instance diagram

[0107]

[0108]

[0109] S3, through the "one-to-two" and internal "one-to-many" touch of the Gestalt piece (Partern-Based One-to-Two Mapping, POTM) strategy, the construction factor of the substantive pattern mode is mapped to the pixel symbiotic feature modeling process as supervision information, and a pixel-level pattern feature coupling extraction model is constructed, as follows:

[0110] (1) Based on the mapping of the tower piece, the classic gray level co-occurrence matrix operator is matched, the parameters of the co-occurrence features such as panoramic window size, direction, statistical distance, and orientation are cooperatively optimized and modeled, the internal observation window size and direction are refined, and the statistical method is optimized, and the whole process of Gestalt pattern embedding is realized. Specifically, let f(x,y) be a two-dimensional digital image, for a pixel Pix(x,y), in the explicit relationship module of the tower piece, an accurate observation window Window(x,y) i is established for each category; in the internal multi-layer implicit relationship module of the tower piece, in each observation window, the tower piece axis point is used for statistical mode fixed-point optimization, that is, the co-occurrence statistical distance, the anisotropy parameter, and the distance between the tower piece axis point, the direction weight are matched, and the coupling statistics of multiple co-occurrence kernels are established. This embodiment refers to this tower piece and GLCM full-parameter POTM mapping scheme as a tower co-occurrence matrix (Gestalt-factors co-occurrence matrix, GFCM), and the calculation strategy is shown in Table 7:

[0111] Table 7 GFCM pattern feature modeling strategy based on Gestalt pattern

[0112]

[0113] For any point (x, y) in the image, find another point (x+dx, y+dy) which is deviated from it by a distance D and the angle G between the coordinate horizontal axis is G, dx<S1, dy<S2, and the gray value of the point pair is (t1, t2). Let the point (x, y) move on the window image (S, C), then various (t1, t2) values are obtained, and the combination of (t1, t2) has L 2 For each window image, the number of occurrences of each (t1, t2) value is counted, and then arranged into a joint probability matrix, which is expressed as: p(t1,t2|D,G)={[(x,y),(x+dx,y+dy)]∈(S1*S2),C|f(x,y)=t1,f(x+dx,y+dy)=t2}.

[0114] By adjusting D and G, the gray co-occurrence matrix p(t1,t2|D,G) of various distances and angles in the window image can be obtained, and then the GFCM of the global image is calculated by using the sliding window.

[0115] In other embodiments, the category i is "internal street road", and the category j "building" belongs to the parallel coexistence relationship. The tile construction factor {S=f(si,di), C=f(ci,gi)} corresponds to the observation window scale and the direction of the community scene to which the category i belongs, which is used for the construction of the window set in the GFCM; the construction factor {D=f(di), G=f(gi)} corresponds to the coexistence distance and the direction between the internal street road and the building, which is used for the construction of the statistical method in the GFCM. As shown in Figure 4 , the peripheral rectangular window is (S, C), which is the observation window; the internal window is (D, G), which is the distance between categories.

[0116] (2) Based on the category i, the two tiles located at the tower base and the tower top are mapped, the GFCM statistical result is obtained, and the average performance of the category i image feature is obtained by summing the accumulation of these statistical matrices, which is expressed as:

[0117] GFCM(S,C,D,G,V)i=∑ G (dG×dD) windows(S,C) ×V;

[0118] Wherein, V=f(SV,CV,α,β) is the confidence of the scene, which is used for the correction of the GFCM statistical result.

[0119] In addition, the scene confidence V is also used to improve the calculation of the traditional four classic GLCM descriptors energy, contrast, entropy and correlation to obtain new GFCM {E, I, H, J}, as follows:

[0120] E(D,G,V)=∑t1 ∑ t2 [p(t1,t2|D,G)×V] 2 ;

[0121] I(D,G,V)=∑ t1 ∑ t2 (t1-t2) 2 [p(t1,t2|D,G)×V];

[0122] H(D,G,V)=-∑ t1 ∑ t2 [p(t1,t2|D,G)×V]log(p(t1,t2|D,G)×V);

[0123]

[0124] where,

[0125] μ t1 =∑ t1 t1∑ t2 p(t1,t2|D,G);

[0126] μ t2 =∑ t2 t2∑ t1 p(t1,t2|D,G);

[0127] σ t1 2 =∑ t1 (t1-μ t1 ) 2 ∑ t2 p(t1,t2|D,G);

[0128] σ t2 2 =∑ t2 (t2-μ t2 ) 2 ∑ t1 p(t1,t2|D,G)。

[0129] The above process, the pattern feature of each pixel is coupled to calculate a total of 4x2xNx(N+M-1)dimension, far lower than the abstract feature dimension of depth convolution calculation, and higher accuracy.

[0130] Through the whole-process linkage's tower piece factor mapping computer mechanism, with multi-core symbiotic calculation and fuzzy calculation mechanism, realize based on pixel pattern feature coupling extraction, not only change traditional based on visual semantic feature redundancy and error irreversible phenomenon, and really cross thematic information and image pattern feature between'semantic gap ', through more intelligent, more light panoramic pattern feature extraction paradigm, realize to the ubiquitous image thematic analysis task accurate understanding and fast response.

[0131] Therefore, the application adopts the above-mentioned thematic image panoramic pattern feature description and extraction method, establishes a cascading transmission and cooperative work framework of real pattern mode, constructs a direct link between high-level symbiotic relationship information and pixel features, truly crosses the semantic gap, realizes the true extraction of thematic pattern features, completes the internal assumption transformation of the pattern analysis model from data-driven to thematic-driven, conforms to the logical information transmission process of human cognitive psychology perception, and provides a new solution for batch high-resolution image pattern analysis.

[0132] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can refer to the practical application of the embodiments of the present application for modification or equivalent replacement, and these modifications or equivalent replacements also cannot make the modified technical solutions deviate from the spirit, framework idea and scope of the technical solutions of the present application.

Claims

1. A topic-oriented image panoramic map feature description and extraction method, characterized in that, The method comprises the following steps: S1, defining a panoramic type concept understanding general model GO based on ontology theory according to the criteria of complete and mutually exclusive, and standardizing the expression as GO={SC, SP}, wherein SC represents a thematic full-combination category system, and SP represents a panoramic type concept based on symbiotic relationship; S2, describing the panoramic type concept understanding general model GO according to the Gestalt theory, and constructing a panoramic type sample library, a symbiotic relationship dependent factor library and a Gestalt pattern library through panoramic sample labeling, sample unit active calculation and Gestalt construction; In the symbiotic relationship dependent factor library, the in-class symbiotic operator includes size S and direction C: the size S represents the average value of the short side of the circumscribed rectangle of the same category, and the direction C represents the average value of the main direction of the same category; the inter-class symbiotic operator includes frequency distance D and frequency orientation G: the distance D represents the minimum value of the high-frequency distance between different categories, and the orientation G represents the direction corresponding to the high-frequency distance between different categories; a scene is expressed by a set of four-dimensional attribute vectors of size S, direction C, distance D and orientation G, and a scene similarity coefficient is designed by using the confidence value of the four-dimensional attribute vector, which is used to measure the similarity of two scenes; In the Gestalt pattern library, according to the panoramic type concept understanding general model, a tile is established for each scene, the tile is provided with an equidistant grid inside, and the marking points on the grid represent the symbiotic relationship; the tile includes size, direction, point distance, point diameter and thickness attribute dimensions; S3, according to the one-to-two and one-to-many touch control strategies of the tile, the construction factors in the Gestalt pattern library are mapped to the pixel symbiotic feature modeling process as supervision information to construct a pixel-level type feature coupling extraction model.

2. The method according to claim 1, wherein, In S1, the thematic full-combination category system includes a scene, a basic class and an accessory class; wherein the basic class is a substantive object defined according to a thematic task, and the accessory class is another substantive object existing in dependence on the basic class; and the inter-class relationship includes parallel symbiotic relationship and parent-child symbiotic relationship; The panoramic type concept based on symbiotic relationship includes four symbiotic spatial relationships of apart symbiosis, adjacent symbiosis, containing symbiosis and complex symbiosis.

3. The method for describing and extracting thematic panoramic image features according to claim 2, characterized in that, The extraction steps of the panoramic type concept based on symbiotic relationship include: According to the basic spatial relationship attribute topology Q, orientation H, metric O and geometric shape R, four-dimensional attributes of a generalized region connection calculus GRCC are formed; then, a grouping superposition strategy is adopted to perform type concept combination deduction and understanding, and topological and geometric shape relationship combination and orientation and metric relationship combination are extracted.

4. The method of claim 1, wherein the method is a topic-oriented image panoramic map feature description and extraction method. In S2, the construction of the panoramic type sample library includes: According to the Gestalt whole-part theory, taking a scene as a carrier of a type unit, a whole sampling strategy based on a scene unit is adopted to label a panoramic sample. The sources of the panoramic sample include the latest global map polygons based on a GIS thematic historical database, global map polygons based on thematic subject interpretation, and global map polygons integrated based on GIS thematic historical data and remote sensing image data interpretation results.

5. The method of claim 1, wherein the method is characterized by: In S2, the construction of the symbiotic relationship dependent factor library includes: First, the shape features of the panoramic sample patches are calculated: the width W and the rotation angle θ of the outer rectangle of the patch, which are used to describe the regular shape of the patch; the fractal dimension F and the rectangular degree Rect, which are used to describe the confidence of the regular shape of the patch. d , wherein F d = 2ln(P / k) / ln(A), P, A are the perimeter, area of the spot, respectively; Rect = A / A MER , A MER is the area of the minimum circumscribed rectangle of the spot. Then, according to the definition of the generalized region connection calculus, the orientation H and the metric O are expressed as the graph factor of the core, and the confidence probability is added to represent the relative fuzzy symbiotic relationship according to the strength of the graph features: The intra-class symbiotic operator represents the shape similarity, including size S and orientation C, which are expressed by the average value of the regular shape of the graph patch, as follows: S i = averge(W); In the formula, subscript i represents the parameters corresponding to class i, and n represents the number of categories of the class. The inter-class symbiotic operator adopts a double histogram threshold analysis method, including a statistical metric histogram, taking the minimum value of the distance histogram peak value between different categories as the distance D, and taking the direction corresponding to the distance D as the orientation G, as follows: D ij = min(Peak(Freq(Dis ij ))1,…,Peak(Freq(Dis ij )) k ); G ij = Dir(max(Num(D ij ) B |B = [45°, 90°, 135°, 180°])); where D ij represents the minimum value of the distance histogram peak between class i and class j, i.e. the minimum value of the high frequency distance; Dis ij represents the distance between the corresponding patches of class i and class j; k represents the number of distance histogram peaks between class i and class j; G ij represents the main direction corresponding to the minimum value of the minimum high frequency distance between classes; Freq represents the frequency histogram statistical function; Peak represents the peak statistical function in the frequency histogram, Num represents the patch number function, (D ij ) B represents D ij patches distributed in direction B, Dir represents the direction function corresponding to the statistical result.

6. The method of claim 1, wherein the method is a topic-oriented image panoramic map feature description and extraction method, characterized in that, Each tower piece represents the full combination of parallel symbiotic relationship or full combination of parent-child symbiotic relationship; the full combination of parallel symbiotic relationship is located at the tower base position, the internal network points marked have symmetry, only half of the symbiotic relationship network points with practical significance and independence are retained; the full combination of parent-child symbiotic relationship is located at the tower top position; each tower piece is staggered and overlapped according to the category system and the size of the range, forming a panoramic format tower of several symbiotic scene cascades.

7. The method for describing and extracting thematic panoramic image features according to claim 6, characterized in that, In the full combination parallel symbiosis, the size of the tile is S=S i +max(S j +D ij ), the orientation of the tile is C=(C i +max(C j +G ij )) / 2, the point distance of the tile is D=D ij , the point diameter of the tile is G=G ij , and the thickness of the tile is V=α×β×(1+SV×CV); the symbiotic mode includes apart symbiosis and adjacent symbiosis; Among them, wherein a represents the weight proportion of the number of pairs of category i and category j appearing in the statistical scale of distance D to the total number of high-frequency distances, Dis ij represents the distance between the corresponding patches of category i and category j, β represents the weight proportion of the number of pairs of category i and category j appearing in the statistical scale of distance D and direction G to the number of directions when the distance is D, SV represents the standard degree of size S, represents the average size of all categories, and CV represents the standard degree of orientation C. In the full combination parent-child symbiosis relationship, the size of the tile is S=S i , the orientation of the tile is C=C i , the point distance of the tile is D=D ij , the point diameter of the tile is G=G ij , the thickness of the tile is V=SV×CV×(1+α×β); the symbiotic mode includes symbiosis, composite symbiosis.

8. The method according to claim 7, wherein, Each element of the scene is composed of (SV, CV, α, β), which represents the confidence probability of the corresponding attribute value appearing in the scene, and the expression of the scene similarity coefficient is: In the formula, V(i, j) represents the scene similarity coefficient of scene Ai and scene Aj, that is, the scene confidence; F represents the sum of the products of the scene attributes.

9. The method for describing and extracting thematic panoramic image features according to claim 8, characterized in that, In S3, the pixel-level graph feature coupling extraction model includes the correction of the statistical results based on the scene confidence V=f(SV, CV, α, β), and the improvement of the four traditional classic descriptors of energy E, contrast I, entropy H and correlation J, as follows: GFCM(S, C, D, G, V) i =∑ G (dG×dD) windows(S,C) ×V; E(D, G, V) = ∑ t1 ∑ t2 [p(t1, t2 | D, G) x V] 2 ; I(D,G,V) = ∑ t1 ∑ t2 (t1-t2) 2 [p(t1,t2|D,G) x V]; H(D, G, V) = -∑ t1 ∑ t2 [p(t1,t2|D,G) x V] log(p(t1,t2|D,G) x V); Among them, p(t1, t2|D, G) = {[(x, y), (x+dx, y+dy)] ∈ (S1*S2), C|f(x, y)=t1, f(x+dx, y+dy)=t2}; μ t1 =∑ t1 t1∑ t2 p(t1,t2|D,G); μ t2 =∑ t2 t2∑ t1 p(t1,t2|D,G); σ t1 2 =∑ t1 (t1-μ t1 ) 2 ∑ t2 p(t1,t2|D,G); σ t2 2 =∑ t2 (t2-μ t2 ) 2 ∑ t1 p(t1,t2|D,G); GFCM(S, C, D, G, V) = ∑i∑j∑t1∑t2p(t1, t2 | D, G) Window(S, C) (t1, t2) (1) i GFCM(S, C, D, G, V) = ∑i∑j∑t1∑t2p(t1, t2 | D, G) Window(S, C) (t1, t2) (1) where GFCM(S, C, D, G, V) denotes the modified statistical result of the gray level co-occurrence matrix obtained by the tiling mapping on any pixel, subscript i denotes the corresponding class; (t1, t2) denotes the gray level values of the point pair corresponding to the distance D and the direction G; Window(S, C) denotes the observation window; (x, y) denotes the position coordinates of any point in the image; (x+dx, y+dy) denotes the point corresponding to the point (x, y) with the distance D and the coordinate horizontal axis angle G; dG and dD respectively denote the statistical direction and the statistical distance of the gray level co-occurrence matrix; p(t1, t2 | D, G) is the joint probability matrix.

Images