A feature map-based target recognition method and device

Abstract

The application discloses a target recognition method and device based on a feature map, and belongs to the technical field of signal processing. The method comprises the following steps: obtaining spectral data of a target to be recognized, performing multi-scale spatial transformation on the spectral data, and constructing a feature map containing arch structures; extracting the vertexes of each arch structure from the feature map, taking the corresponding wave number and scale as initial features, and determining the maximum scale of the target to be recognized; for each reference feature in a reference feature library, correcting the scale of the initial feature based on the maximum scale of the reference feature and the maximum scale of the target to be recognized, and obtaining a corrected feature; matching the corrected feature with each reference feature, calculating a matching cost, selecting the reference feature with the minimum matching cost as a candidate feature, and combining a preset discrimination threshold to determine the recognition result of the target to be recognized. Therefore, by implementing the application, the problem of inaccurate target recognition in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and specifically to a target recognition method and apparatus based on feature maps. Background Technology

[0002] Target identification technology based on spectral data has significant application value in fields such as remote sensing, material analysis, and security monitoring. By acquiring the spectral data of the target to be identified and analyzing the spectral intensity distribution patterns at different wavenumbers, it is possible to effectively distinguish target objects with different properties. Due to the differences in the physicochemical properties of different substances, they will exhibit distinctive spectral intensity peaks or absorption characteristics at specific wavenumber positions. Utilizing these spectral characteristics, effective target information can be extracted from complex background interference, thereby achieving target detection and identification.

[0003] However, existing identification methods are often inaccurate when using spectral data for target identification. These methods face the problem that spectral features are unstable as environmental and physical conditions change. Due to fluctuations in temperature, pressure, or the target's own state in the acquisition environment, the target's spectral curve often undergoes nonlinear broadening or scaling, resulting in spectral feature deformation at different scales. However, most existing identification methods directly extract local features based on the original spectral intensity for comparison, lacking a mechanism to place spectral data in a multi-scale space to explore its structural evolution. At the same time, they cannot effectively correct for scale inconsistencies between the target and the reference template caused by environmental differences, thus leading to inaccurate target identification. Summary of the Invention

[0004] This invention provides a target recognition method and apparatus based on feature maps, which can solve the problem of insufficient target recognition in the prior art.

[0005] An embodiment of the present invention provides a target recognition method based on feature map, comprising: Acquire spectral data of the target to be identified; wherein, the spectral data is used to characterize the spectral intensity of the target to be identified at different wavenumbers; A multi-scale spatial transformation is performed on the spectral data of the target to be identified to generate a feature map of the target; wherein, the feature map contains several arched structures used to characterize the spectrum as it changes with scale. The vertices of each arch structure are extracted from the feature map; the wavenumber and scale of each arch structure vertex in the feature map are determined as the initial features of the target to be identified; the maximum scale of each arch structure vertex in the feature map is taken as the maximum scale of the target to be identified. For each reference feature in the preset reference feature library, the maximum scale of the current reference feature is determined according to the scale of the vertices of each arch structure corresponding to the current reference feature; based on the maximum scale of the current reference feature and the maximum scale of the target to be identified, the scale of the initial feature of the target to be identified is corrected to generate the corrected feature of the target to be identified; the corrected feature of the target to be identified is matched with the current reference feature to generate the matching cost of the current reference feature; wherein, the reference feature includes the wave number and scale of the vertices of each arch structure corresponding to the reference target; From the matching costs of each reference feature, the reference feature with the lowest matching cost is selected as the candidate feature; based on the candidate feature and the preset discrimination threshold, the recognition result of the target to be identified is determined.

[0006] Furthermore, multi-scale spatial transformation is performed on the spectral data of the target to be identified to generate a feature map of the target, including: Based on a preset Gaussian kernel function, convolution operations are performed on the spectral data of the target to be identified at preset scales to generate convolution results corresponding to each scale. For each scale, calculate the second derivative of the convolution result with respect to the wavenumber corresponding to the current scale to generate the current second derivative curve; extract the zero-crossing points of the current second derivative curve from the current second derivative curve as the zero-crossing points corresponding to the current scale. A two-dimensional coordinate system is constructed with wavenumber as the abscissa and scale as the ordinate; The zero-crossing points corresponding to each scale are mapped to a two-dimensional coordinate system to generate a feature map of the target to be identified.

[0007] Furthermore, the vertices of each arch structure are extracted from the feature map, including: Contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing point contour lines; The outlines of each zero-crossing point are defined as the arched structures; For each arch structure, select the zero-crossing point with the largest ordinate from the zero-crossing points corresponding to the current arch structure, and use it as the vertex of the current arch structure.

[0008] Furthermore, contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing contour lines, including: Repeat the contour extraction operation until all zero-crossing points in the feature map are marked, generating several closed zero-crossing contour lines; The contour extraction operation includes: Based on the preset traversal order, the feature map is scanned, and the first zero-crossing point that has not been marked as processed is taken as the current starting node. Based on the current starting node, determine the current search path; add the current search path to the search path set, and mark all zero-crossing points on the current search path as processed; Determine if there are any unmarked zero-crossing points in the feature map. If not, perform a closure check on each search path in the search path set to determine the contour lines of each closed zero-crossing point. If so, scan the feature map based on the preset traversal order and update the first unmarked zero-crossing point found to the current starting node.

[0009] Furthermore, based on the current starting node, the current search path is determined, including: Add the current starting node to the current connection path; Repeat the adjacent zero-crossing point connection operation until there are no zero-crossing points that have not been added to the current connection path and have not been marked as processed, and generate the current search path; The adjacent zero-crossing point connection operation includes: Within the preset neighborhood of the current connected node, search for any zero-crossing points that have not been added to the current connected path and have not been marked as processed; If such a zero-crossing point exists, it will be selected as a candidate zero-crossing point within the preset neighborhood of the current connected node that has not been added to the current connected path and has not been marked as processed; where the initial connected node is the current starting node. Based on the candidate zero-crossing points, construct a set of neighborhood zero-crossing points; Based on the preset direction search priority, a unique target zero-crossing point is determined from the set of neighboring zero-crossing points; the target zero-crossing point is added to the current connection path, and the target zero-crossing point is updated as the current connection node; If it does not exist, the current connection path will be used as the current search path.

[0010] Furthermore, a closure check is performed on each search path in the search path set to determine the zero-crossing contour lines of each closed path, including: For each search path in the set of search paths, determine the start and end nodes of the current search path; Determine whether the starting node of the current search path is within the preset neighborhood of the ending node of the current search path; If so, then determine that the current search path satisfies the closure condition, and define the current search path as a closed zero-crossing contour line; If not, then the current search path does not meet the closure condition; If none of the search paths meet the closure condition, calculate the confidence level of each search path based on the start and end nodes of the search path; sort the search paths in descending order based on the confidence level to generate a path sorting result; select the first preset number of search paths from the path sorting result as each path to be compensated; perform closure connection processing on each path to be compensated to generate a closed zero-crossing contour line.

[0011] Furthermore, based on the maximum scale of the current reference features and the maximum scale of the target to be identified, the scale of the initial features of the target to be identified is corrected to generate corrected features of the target to be identified, including: Calculate the ratio of the maximum scale of the current reference feature to the maximum scale of the target to be identified, and generate the current scale correction coefficient; Based on the current scale correction coefficient, the scale of each vertex in the initial features of the target to be identified is corrected to generate the corrected scale of each vertex in the initial features of the target to be identified. The wavenumbers of each vertex in the initial features of the target to be identified are combined with the correction scale to generate the corrected features of the target to be identified.

[0012] Furthermore, the corrected features of the target to be identified are matched with the current reference features to generate the matching cost of the current reference features, including: Based on the wave number and correction scale of each vertex in the modified features of the target to be identified, construct the feature vector to be identified; Construct the current reference feature vector based on the wave number and scale of each vertex in the current reference feature; Calculate the Euclidean distance between the feature vector to be identified and the current reference feature vector, and generate the matching cost of the current reference feature.

[0013] Furthermore, based on candidate features and a preset discrimination threshold, the recognition result of the target to be identified is determined, including: Determine whether the matching cost of the candidate feature is less than the preset discrimination threshold. If so, use the candidate feature as the recognition result of the target to be identified. If not, then the identification result of the target to be identified is determined to be an unknown target.

[0014] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0015] One embodiment of the present invention provides a target recognition device based on feature map, comprising: a data acquisition module, a multi-scale spatial transformation module, a feature correction module, and a recognition and determination module; The data acquisition module is used to acquire the spectral data of the target to be identified; wherein, the spectral data is used to characterize the spectral intensity of the target to be identified at different wavenumbers; The multi-scale spatial transformation module performs multi-scale spatial transformation on the spectral data of the target to be identified, generating a feature map of the target; wherein, the feature map contains several arched structures used to characterize the spectrum as it changes with scale. The feature correction module is used to extract the vertices of each arch structure from the feature map; determine the wavenumber and scale of each arch structure vertex in the feature map as the initial features of the target to be identified; take the maximum scale of each arch structure vertex in the feature map as the maximum scale of the target to be identified; for each reference feature in the preset reference feature library, determine the maximum scale of the current reference feature according to the scale of each arch structure vertex corresponding to the current reference feature; based on the maximum scale of the current reference feature and the maximum scale of the target to be identified, correct the scale of the initial feature of the target to be identified to generate the corrected feature of the target to be identified; perform matching calculation between the corrected feature of the target to be identified and the current reference feature to generate the matching cost of the current reference feature; wherein, the reference feature includes the wavenumber and scale of each arch structure vertex corresponding to the reference target; The identification and determination module is used to select the reference feature with the smallest matching cost from the matching costs of each reference feature as a candidate feature; and to determine the identification result of the target to be identified based on the candidate feature and a preset discrimination threshold.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a target recognition method, apparatus, electronic device, and storage medium based on feature maps. The method involves: acquiring spectral data of the target to be identified; performing multi-scale spatial transformation on the spectral data to construct a feature map containing arched structures; extracting the vertices of each arched structure from the feature map and using their corresponding wavenumbers and scales as initial features of the target to be identified, while simultaneously determining the maximum scale of the target; for each reference feature in a reference feature library, determining the maximum scale of the reference feature based on the scale of the arched structure vertices of the reference feature, and correcting the scale of the initial target feature based on the maximum scale of the reference feature and the maximum scale of the target feature to obtain corrected features; matching the corrected features with each reference feature, calculating the matching cost, selecting the reference feature with the lowest matching cost as a candidate feature, and determining the identification result of the target to be identified by combining it with a preset discrimination threshold.

[0017] This invention effectively addresses the problem of existing technologies lacking the mechanism for mining the evolution of spectral structures in a multi-scale space by performing multi-scale spatial transformations on the spectral data of the target to be identified and generating a feature map containing an arched structure. Furthermore, by utilizing the maximum scale of the target and reference features to perform scale correction on the initial features of the target, it overcomes the technical deficiency of existing methods that cannot effectively correct for scale inconsistencies caused by environmental and physical state fluctuations. This eliminates matching errors caused by spectral nonlinear broadening or scaling, significantly improving the accuracy of target identification. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a target recognition method based on feature maps according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of a target recognition device based on feature map according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, to address the problem of inaccurate target recognition in existing technologies, an embodiment of the present invention provides a target recognition method based on feature maps, comprising at least the following steps: Step S1: Obtain the spectral data of the target to be identified; wherein the spectral data is used to characterize the spectral intensity of the target to be identified at different wavenumbers.

[0022] Specifically, the spectral data of the target to be identified is a discrete data sequence formed by detecting the target with a spectral detection device and then digitally sampling it. Mathematically, the spectral data of the target is represented as one-dimensional vector data. It consists of a series of data points arranged in wavenumber order, each data point containing a wavenumber value and its corresponding spectral intensity value. The wavenumber, as the independent variable, identifies the specific position of the spectrum in the frequency domain, while the spectral intensity, as the dependent variable, numerically describes the amplitude of radiant or reflected energy of the target at a specific wavenumber position.

[0023] In this embodiment, the target to be identified specifically refers to an observation object with a specific chemical composition or specific material structure within a preset spectral detection band, and the spectral response curve of the target to be identified possesses identifiable peak and trough characteristics. Targets to be identified specifically include chemical gas clouds with specific infrared radiation characteristics, artificial camouflage objects coated with specific spectral reflective materials, and geographical surfaces with specific mineral compositions. Due to the influence of their own temperature, environmental pressure, or material concentration, the spectral intensity distribution of these targets will undergo nonlinear scale-scaling deformation while retaining the peak and trough topology. This spectral characteristic, possessing structural stability but lacking scale stability, is the physical basis for the multi-scale spatial transformation and maximum scale correction techniques used in subsequent steps of this application.

[0024] In actual detection scenarios, the target to be identified is often surrounded by complex background clutter, or is obscured by interference and camouflage, resulting in the target's spectral data exhibiting weak signal strength and low signal-to-noise ratio and signal-to-clutter ratio.

[0025] By acquiring the spectral data of the target to be identified through step S1, the unique spectral response of the target in a specific spectral band can be converted into a digital signal that can be processed by a computer. This ensures that the subsequent processing steps are based on the essential physical characteristics of the target and provides accurate raw data support for the subsequent use of multi-scale spatial transformation to mine the scale-dependent structural evolution laws contained in the spectral data.

[0026] Step S2: Perform multi-scale spatial transformation on the spectral data of the target to be identified to generate a feature map of the target; wherein the feature map contains several arched structures used to characterize the spectrum as it changes with scale.

[0027] In a preferred embodiment, a multi-scale spatial transformation is performed on the spectral data of the target to be identified to generate a feature map of the target, including: Based on a preset Gaussian kernel function, convolution operations are performed on the spectral data of the target to be identified at preset scales to generate convolution results corresponding to each scale. For each scale, calculate the second derivative of the convolution result with respect to the wavenumber corresponding to the current scale to generate the current second derivative curve; extract the zero-crossing points of the current second derivative curve from the current second derivative curve as the zero-crossing points corresponding to the current scale. A two-dimensional coordinate system is constructed with wavenumber as the abscissa and scale as the ordinate; The zero-crossing points corresponding to each scale are mapped to a two-dimensional coordinate system to generate a feature map of the target to be identified.

[0028] Specifically, multi-scale spatial transformation of the spectral data of the target to be identified is based on the one-dimensional curvature scale space theory. The core idea of ​​multi-scale spatial transformation is to treat the one-dimensional spectral signal of the target to be identified as a planar curve, and use Gaussian functions with different scales as smoothing filters to continuously smooth the spectral data of the target to be identified, thereby simulating the evolution of the spectral features of the target to be identified at different observation scales.

[0029] In a specific embodiment, the specific operation process of performing multi-scale spatial transformation on the spectral data of the target to be identified to generate the feature map of the target is as follows: First, based on a preset Gaussian kernel function, convolution operations are performed on the spectral data of the target to be identified at preset scales, generating convolution results for each scale. The spectral data of the target to be identified is mathematically represented as a spectral intensity function with respect to wavenumber. ,in Represents wavenumber. Multiscale spatial transformation introduces scale. To construct a one-dimensional Gaussian kernel function One-dimensional Gaussian kernel function The mathematical expression is as follows: in, Represents the wavenumber variable. Indicates scale. Represents pi (π). Represented by natural constant An exponential function with base 0.

[0030] Based on the aforementioned one-dimensional Gaussian kernel function, the spectral data of the target to be identified... Perform convolution operations to generate convolution results for each scale. The mathematical expression for convolution is as follows: in, Indicated at a scale of The convolution result after smoothing the spectral data of the target to be identified. Represents the integral variable. This represents the convolution operator. In actual digital signal processing, the above integration operation is implemented through discrete convolution summation. The preset scales refer to starting from a small initial scale and gradually increasing the scale value according to a preset step size until a preset termination condition is met. As the scale... The increase of , the convolution result The corresponding high-frequency noise components are gradually filtered out, and the spectral profile of the target to be identified gradually becomes smoother.

[0031] Secondly, for each scale, the second derivative of the convolution result with respect to the wavenumber is calculated to generate the current second derivative curve. From this curve, the zero-crossing points are extracted and used as the zero-crossing points corresponding to the current scale. (Convolution result) The inflection point corresponds to the boundary between the concavity and convexity of the spectral curve, and mathematically, the inflection point corresponds to the position where the second derivative is zero. Therefore, the calculation of the convolution result... Regarding wavenumber The second derivative Solve the following equation to determine the location of the zero-crossing point: For each specific scale The solution to the above equation is the zero-crossing point corresponding to the current scale. The position of the zero-crossing point on the wavenumber axis reflects the structural inflection point of the spectral curve of the target to be identified at that scale.

[0032] Finally, a two-dimensional coordinate system is constructed with wavenumber as the x-axis and scale as the y-axis. The zero-crossing points corresponding to each scale are mapped onto the two-dimensional coordinate system to generate a feature map of the target to be identified. In the two-dimensional coordinate system, each recorded zero-crossing point is represented by a coordinate pair. It means that, among them The position of the wavenumber at the zero-crossing point. This corresponds to the scale value. As the scale... As the spectral curve of the target object undergoes continuous change from small to large, the minute bumps and depressions on the curve are gradually smoothed out, causing the corresponding pair of zero-crossing points to approach each other on the wavenumber axis and eventually converge. Marking and connecting these zero-crossing points at different scales in a two-dimensional coordinate system forms several arched structures. Specifically, in the constructed two-dimensional coordinate system, the horizontal axis represents the wavenumber, and the vertical axis represents the scale of the smoothing process. Geometrically, the arched structure presents itself as a downward-opening parabolic closed curve. As the scale of the vertical axis increases, the two branch curves at the bottom gradually converge towards the center and intersect at a specific scale height, forming the highest point of the arched structure, which is the apex of the arched structure. The two branch curves at the bottom correspond to the two zero-crossing points on either side of a local spectral peak or trough. Each arched structure represents the lifespan of a spectral peak or valley feature of the target object as it changes with scale, and the apex of the arched structure represents the critical scale at which the spectral feature is completely smoothed out.

[0033] The feature map containing the arched structure generated in step S2 can map the spectral data of the target to be identified from the easily disturbed one-dimensional intensity space to the topologically invariant two-dimensional scale space, effectively eliminating random noise in the spectral data and extracting the robust intrinsic structural features of the target to be identified.

[0034] It should be noted that the preset scales constitute an ordered scale sequence for multi-scale spatial transformation. The determination of the scale sequence follows a strategy of gradually increasing from small scales to large scales according to the principle of equal spacing.

[0035] The specific scale selection process first determines an initial scale. The initial scale is set to a small positive number, typically 1 or close to the minimum spectral resolution of the spectrometer. The initial scale selection must ensure that as many fine texture features and weak signal features as possible are preserved in the spectral data of the target to be identified during the initial stage of multi-scale transformation, preventing highly recognizable narrow-band spectral features from being prematurely smoothed out at lower scales. Next, the initial scale is iteratively accumulated according to a preset scale step size to generate each subsequent scale. The scale step size is typically set to 0.5 or less than the initial scale value. The rationale for selecting the scale step size is to ensure that the generated arched structure in the feature map has sufficient smoothness and connectivity. Only sufficiently dense scale sampling points can ensure that discrete zero-crossing points at different scales are connected into a continuous and closed contour curve, avoiding the situation where zero-crossing points on the same arched structure are broken or cannot be correctly correlated due to excessively large scale sampling intervals. The incremental scaling process continues until a preset termination condition is met. The termination condition is set to the number of zero-crossing points of the second derivative calculated at the current scale, which means that the smoothed spectral curve has evolved into a monotonic form or only retains the overall outline without local inflection points, and no longer has high-frequency detail features.

[0036] By adopting the scale sequence selection strategy based on the initial scale, scale step size and termination condition, it is possible to effectively suppress background random noise interference and construct a complete and continuous feature map that represents the entire process of the spectral features of the target to be identified from generation to disappearance. This provides a high-precision structured data foundation for the subsequent accurate extraction of the maximum scale parameter and initial features.

[0037] Step S3: Extract the vertices of each arch structure from the feature map; determine the wavenumber and scale of each arch structure vertex in the feature map as the initial features of the target to be identified; take the maximum scale of each arch structure vertex in the feature map as the maximum scale of the target to be identified.

[0038] In a preferred embodiment, the vertices of each arch structure are extracted from the feature map, including: Contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing point contour lines; The outlines of each zero-crossing point are defined as the arched structures; For each arch structure, select the zero-crossing point with the largest ordinate from the zero-crossing points corresponding to the current arch structure, and use it as the vertex of the current arch structure.

[0039] In a preferred embodiment, contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing point contour lines, including: Repeat the contour extraction operation until all zero-crossing points in the feature map are marked, generating several closed zero-crossing contour lines; The contour extraction operation includes: Based on the preset traversal order, the feature map is scanned, and the first zero-crossing point that has not been marked as processed is taken as the current starting node. Based on the current starting node, determine the current search path; add the current search path to the search path set, and mark all zero-crossing points on the current search path as processed; Determine if there are any unmarked zero-crossing points in the feature map. If not, perform a closure check on each search path in the search path set to determine the contour lines of each closed zero-crossing point. If so, scan the feature map based on the preset traversal order and update the first unmarked zero-crossing point found to the current starting node.

[0040] In a preferred embodiment, determining the current search path based on the current starting node includes: Add the current starting node to the current connection path; Repeat the adjacent zero-crossing point connection operation until there are no zero-crossing points that have not been added to the current connection path and have not been marked as processed, and generate the current search path; The adjacent zero-crossing point connection operation includes: Within the preset neighborhood of the current connected node, search for any zero-crossing points that have not been added to the current connected path and have not been marked as processed; If such a zero-crossing point exists, it will be selected as a candidate zero-crossing point within the preset neighborhood of the current connected node that has not been added to the current connected path and has not been marked as processed; where the initial connected node is the current starting node. Based on the candidate zero-crossing points, construct a set of neighborhood zero-crossing points; Based on the preset direction search priority, a unique target zero-crossing point is determined from the set of neighboring zero-crossing points; the target zero-crossing point is added to the current connection path, and the target zero-crossing point is updated as the current connection node; If it does not exist, the current connection path will be used as the current search path.

[0041] In a preferred embodiment, a closure check is performed on each search path in the search path set to determine the zero-crossing contour line of each closed path, including: For each search path in the set of search paths, determine the start and end nodes of the current search path; Determine whether the starting node of the current search path is within the preset neighborhood of the ending node of the current search path; If so, then determine that the current search path satisfies the closure condition, and define the current search path as a closed zero-crossing contour line; If not, then the current search path does not meet the closure condition; If none of the search paths meet the closure condition, calculate the confidence level of each search path based on the start and end nodes of the search path; sort the search paths in descending order based on the confidence level to generate a path sorting result; select the first preset number of search paths from the path sorting result as each path to be compensated; perform closure connection processing on each path to be compensated to generate a closed zero-crossing contour line.

[0042] Specifically, the feature map of the target to be identified is generated by multi-scale spatial transformation. Essentially, the feature map is a two-dimensional binary image composed of zero-crossing points of the second derivative. Since spectral features undergo a process of gradual disappearance with increasing scale, these zero-crossing points of the second derivative exhibit a tendency to gradually converge and eventually merge in the two-dimensional coordinate system, thus forming several closed curve structures, which are called arch structures. Each arch structure corresponds to the periodicity of a specific spectral fingerprint feature of the target to be identified in the multi-scale space.

[0043] Extracting the vertices of each arch structure from the feature map is a crucial operation for obtaining robust features with scale invariance. The vertices of the arch structures are located at their highest points, corresponding to the critical scale at which the spectral features disappear. Determining the wavenumber and scale values ​​of each arch structure's vertices in the feature map as the initial features of the target to be identified condenses the high-dimensional spectral curve shape into a sparse and representative set of feature points. Simultaneously, the feature point with the largest scale value is selected from all extracted initial features, and this largest scale value is taken as the maximum scale of the target to be identified. The maximum scale of the target to be identified reflects the lifespan of the most significant and smooth-resistant structural feature in the overall spectral profile of the target. This maximum scale will subsequently be used as a benchmark parameter to correct for scale differences between the target and the reference feature library.

[0044] In a specific embodiment, the vertices of each arch structure are extracted from the feature map. The specific execution process includes: First, contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing contour lines. Since the feature map is generated based on discrete scale steps, there may be local discontinuities or breaks in the zero-crossing points. Contour tracing can connect discrete zero-crossing points belonging to the same spectral feature evolution process into a continuous curve trajectory.

[0045] Secondly, the identified closed zero-crossing contour lines are defined as arched structures. A complete arched structure should be a closed loop, meaning that the spectral features are generated at low scales and eventually close and disappear as the scale increases.

[0046] Finally, for each arch structure, the coordinate data of all zero-crossing points contained in the current arch structure are traversed, the ordinate values ​​of all zero-crossing points are compared, and the zero-crossing point with the largest ordinate is selected as the vertex of the current arch structure. The ordinate corresponds to the scale axis in the feature map, so the zero-crossing point with the largest ordinate is the highest point of the arch structure. The coordinate information of the highest point includes the center wavenumber position of the feature and the disappearance scale of the feature.

[0047] In one specific embodiment, contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing contour lines. The specific execution process includes: The contour extraction operation is repeatedly performed using an iterative scanning method until all zero-crossing points in the feature map are marked as processed, thereby generating several closed zero-crossing contour lines. The specific steps of the contour extraction operation are as follows: Based on a preset traversal order, the feature map is scanned point by point. The preset traversal order is usually set to start from the lower left corner of the feature map and proceed in a raster scan manner from left to right and from bottom to top. The status marker of each zero-crossing point scanned is checked, and the first zero-crossing point that is not marked as processed is taken as the current starting node. The current starting node marks the starting point of a new potential arch structure.

[0048] Based on the current starting node, a path search mechanism is initiated to determine the current search path. The path search mechanism extends along the distribution of zero-crossing points, adding each found zero-crossing point sequentially to the current search path. During the search process, all zero-crossing points on the current search path are marked as processed to prevent repeated extraction of the same arch structure in subsequent scanning steps.

[0049] After completing one path search, it is determined whether there are any unmarked zero-crossings in the feature map. If there are unmarked zero-crossings, it means that there are still unextracted structures in the feature map. Based on the preset traversal order, the feature map is scanned again, and the first unmarked zero-crossing is updated as the current starting node, triggering a new round of contour extraction. If there are no unmarked zero-crossings, it means that all information in the feature map has been traversed. The scanning stops, and a closure check is performed on each search path in the search path set. Closed zero-crossing contour lines are then selected and determined from the search path set.

[0050] In one specific embodiment, the current search path is determined based on the current starting node, and the specific execution process includes: First, the current starting node is used as the initial connection node and added to the current connection path.

[0051] Subsequently, the adjacent zero-crossing connection operation is repeated until a stopping condition is met. The stopping condition is that there are no zero-crossings that have not been added to the current connection path and have not been marked as processed, meaning that the current search path has extended to an endpoint or closure point. After the loop ends, the current search path is generated.

[0052] The specific logic for the adjacent zero-crossing point connection operation is as follows: Within the preset neighborhood of the current connected node, search for any zero-crossing points that have not been added to the current connected path and have not been marked as processed. The preset neighborhood is usually set as an eight-neighbor pixel region centered on the current connected node, that is, including pixels in the eight directions around the current connected node: up, down, left, right, upper left, lower left, upper right, and lower right.

[0053] If the search result finds a zero-crossing point, all zero-crossing points not added to the current connection path and not marked as processed within the preset neighborhood of the current connected node will be considered as candidate zero-crossing points. A neighborhood zero-crossing point set will be constructed based on these candidate zero-crossing points. To ensure the smoothness and directionality of contour tracking and avoid sharp-angle turns or intersections in the path, a unique target zero-crossing point needs to be determined from the neighborhood zero-crossing point set based on a preset direction search priority. The preset direction search priority is set based on the extension direction of the current connection path, prioritizing the candidate zero-crossing point with the smallest angle to the extension direction. The determined target zero-crossing point will be added to the current connection path, and the target zero-crossing point will be updated to the current connected node so that the search can continue with the new connected node as the center in the next iteration.

[0054] If the search result is not found, it means that there are no valid zero-crossing points around the current connection node that can be extended, and the path search has reached its end. At this time, the current connection path is output as the complete current search path.

[0055] In a specific embodiment, a closure check is performed on each search path in the search path set to determine the zero-crossing contour line of each closed path. The specific execution process includes: Due to background noise interference or discrete sampling errors during multi-scale transformation, the extracted search paths may not all be perfectly closed curves. Therefore, closure checks and compensation are required. For each search path in the search path set, extract the coordinates of the starting node and the ending node of the current search path.

[0056] Determine whether the starting node of the current search path is within a preset neighborhood of the ending node of the current search path. This determination is used to detect whether the beginning and end of the search path are connected. The size of the preset neighborhood is set to an allowable closure error range.

[0057] If so, it means that the search path forms a closed loop. Therefore, the current search path is determined to meet the closure condition, and the current search path is directly determined as a closed zero-crossing contour line.

[0058] If not, it indicates that the search path has been broken, and therefore the current search path does not meet the closure condition.

[0059] Under conditions of extreme noise or weak signal, it is possible that none of the search paths meet the closure condition. To avoid the complete loss of effective features, a compensation mechanism needs to be activated: based on the spatial geometric relationship between the start and end nodes of the search path, the confidence score of each search path is calculated. The confidence score is calculated based on the Euclidean distance between the start and end nodes; the smaller the distance value, the closer the search path is to a closed state, and the higher the confidence score.

[0060] Specifically, the confidence score for each search path can be calculated using the following formula: in, Indicates the first Confidence of each search path; Indicates the first The total number of zero-crossing points contained in the search path; the total number of zero-crossing points is used to characterize the length of the path. Indicates the first The Euclidean distance between the starting and ending nodes of a search path.

[0061] The above formula can transform abstract spatial geometric relationships into specific quantifiable values, ensuring that search paths with closer distances between the start and end nodes and richer feature information can obtain higher confidence scores, thus being prioritized as paths to be compensated in subsequent descending sorting.

[0062] The search paths are sorted in descending order based on the calculated confidence scores, generating a path ranking result. From the path ranking result, the top preset number of search paths (e.g., the top N with the highest confidence scores) are selected as the paths to be compensated. Each path to be compensated is then subjected to a closed-loop connection process, artificially connecting the start and end nodes of the path to be compensated using linear interpolation or curve fitting, forcibly generating closed zero-crossing contour lines.

[0063] Through the above step S3, the arched vertex features representing the essential attributes of the target to be identified can be accurately extracted from the feature map that is affected by noise. In particular, the introduction of a confidence-based path compensation mechanism effectively solves the problem of feature map fragmentation caused by environmental interference, which makes it impossible to extract features. This significantly improves the robustness of the target recognition method in low signal-to-noise ratio environments and the completeness of feature extraction.

[0064] For example, to illustrate in detail how a computer identifies closed zero-crossing contours from discrete pixels in a feature map through contour tracing, the following specific numerical processing scenario is set up. Assume the feature map generated in step S2 is stored in a simplified 10x10 two-dimensional logical matrix, with the horizontal coordinate ranging from 0 to 9 and the vertical coordinate ranging from 0 to 9. In the two-dimensional logical matrix, a value of 1 indicates the existence of a zero-crossing point of the second derivative, and a value of 0 indicates the background. The computer is configured with a state matrix of the same size to record whether each pixel has been processed and marked; initially, all positions are unprocessed. Assume there exists a closed rectangular structure in the feature map consisting of four pixels, with coordinates P1(2, 2), P2(2, 3), P3(3, 3), and P4(3, 2).

[0065] The computer scans the feature map point by point according to a preset traversal order. The computer scans the region from coordinates (0, 0) to (2, 1), detecting pixel values ​​of 0, and continues scanning. When it reaches coordinate (2, 2), the computer detects a pixel value of 1 and checks the state matrix to confirm that coordinate (2, 2) is unprocessed. The computer designates coordinate (2, 2) as the current starting node, adds it to the current connection path list, and marks it as processed in the state matrix. At this point, the current connection path is: [(2, 2)].

[0066] The computer initiates a neighborhood search loop centered on the current connected node (2, 2) to determine the subsequent path direction. In the first path extension loop, the computer checks the eight neighbors of node (2, 2). The computer finds that the pixel values ​​at coordinates (2, 3) and (3, 2) are 1 and their state is unprocessed. Based on the preset direction priority, the computer selects coordinates (2, 3) as the target zero-crossing point. The computer adds coordinates (2, 3) to the current connected path and marks coordinates (2, 3) as processed. At this point, the current connected path is updated to: [(2, 2), (2, 3)]. The current connected node is updated to (2, 3). In the second path extension loop, the computer checks the eight neighbors of node (2, 3). The computer finds that the pixel value at coordinates (3, 3) is 1 and its state is unprocessed. The computer adds coordinates (3, 3) to the current connected path and marks coordinates (3, 3) as processed. At this point, the current connection path is updated to: [(2, 2), (2, 3), (3, 3)]. The current connection node is updated to (3, 3). In the third path extension loop, the computer checks the eight neighbors of node (3, 3). The computer finds that the pixel value of coordinate (3, 2) in the neighborhood is 1 and its state is unprocessed. The computer adds coordinate (3, 2) to the current connection path and marks coordinate (3, 2) as processed. At this point, the current connection path is updated to: [(2, 2), (2, 3), (3, 3), (3, 2)]. The current connection node is updated to (3, 2). In the fourth path extension loop, the computer checks the eight neighbors of node (3, 2). The computer finds that the valid pixels (2, 2) and (3, 3) in the neighborhood have both been marked as processed, and there are no unprocessed adjacent zero-crossing points. The computer determines that there are no candidate zero-crossing points, terminates the internal path extension loop, and outputs the current connection path as the current search path.

[0067] The computer obtains the starting node (2, 2) and ending node (3, 2) of the current search path. It calculates the positional relationship between these nodes, determining that the starting node (2, 2) is within the eight-neighborhood of the ending node (3, 2). Based on this, the computer determines that the current search path satisfies the closure condition, defining the path [(2, 2), (2, 3), (3, 3), (3, 2)] as a closed zero-crossing contour line, i.e., the first arch structure. Subsequently, the computer continues the global scan from the previously interrupted coordinates (2, 2). When scanning to the positions (2, 3), (3, 2), and (3, 3), since these positions have already been marked as processed in the state matrix, the computer skips these positions and continues searching for the next unprocessed potential structure until the entire feature map has been traversed.

[0068] Step S4: For each reference feature in the preset reference feature library, determine the maximum scale of the current reference feature based on the scale of the vertices of each arch structure corresponding to the current reference feature; based on the maximum scale of the current reference feature and the maximum scale of the target to be identified, correct the scale of the initial feature of the target to be identified to generate the corrected feature of the target to be identified; perform matching calculation between the corrected feature of the target to be identified and the current reference feature to generate the matching cost of the current reference feature; wherein, the reference feature includes the wave number and scale of the vertices of each arch structure corresponding to the reference target.

[0069] In a preferred embodiment, based on the maximum scale of the current reference features and the maximum scale of the target to be identified, the scale of the initial features of the target to be identified is corrected to generate corrected features of the target to be identified, including: Calculate the ratio of the maximum scale of the current reference feature to the maximum scale of the target to be identified, and generate the current scale correction coefficient; Based on the current scale correction coefficient, the scale of each vertex in the initial features of the target to be identified is corrected to generate the corrected scale of each vertex in the initial features of the target to be identified. The wavenumbers of each vertex in the initial features of the target to be identified are combined with the correction scale to generate the corrected features of the target to be identified.

[0070] In a preferred embodiment, the modified features of the target to be identified are matched with the current reference features to generate a matching cost for the current reference features, including: Based on the wave number and correction scale of each vertex in the modified features of the target to be identified, construct the feature vector to be identified; Construct the current reference feature vector based on the wave number and scale of each vertex in the current reference feature; Calculate the Euclidean distance between the feature vector to be identified and the current reference feature vector, and generate the matching cost of the current reference feature.

[0071] Specifically, the preset reference feature library is a pre-established collection of standard spectral feature data containing various known targets. For each reference feature in the library, the computer traverses the vertex information of all the arched structures contained in the reference feature, identifies the scale with the largest value among the reference features, and marks the scale with the largest value as the maximum scale of the current reference feature. At the same time, the computer obtains the maximum scale of the target to be identified determined in step S3. The difference between the maximum scale of the target to be identified and the maximum scale of the current reference feature objectively reflects the degree of signal scale scaling of the target to be identified relative to the standard reference environment under the current detection environment. By aligning the maximum scale of the target to be identified with the maximum scale of the reference feature, the effects of overall nonlinear spectral scaling caused by changes in detection distance, ambient temperature, or pressure can be eliminated, thereby ensuring that subsequent matching calculations are performed on a unified scale benchmark.

[0072] In one specific embodiment, based on the maximum scale of the current reference features and the maximum scale of the target to be identified, the scale of the initial features of the target to be identified is corrected to generate corrected features of the target to be identified. The specific execution process includes: First, the ratio of the maximum scale of the current reference feature to the maximum scale of the target to be identified is calculated to generate the current scale correction coefficient. The scale correction coefficient is used to quantify the overall scaling difference between the target to be identified and the reference target in scale space. The mathematical expression for calculating the scale correction coefficient is as follows: in, This represents the current scale correction factor. This represents the maximum scale of the current reference feature. This indicates the maximum scale of the target to be identified.

[0073] Secondly, based on the current scale correction coefficient, the scale of each vertex in the initial features of the target to be identified is corrected, generating a corrected scale for each vertex in the initial features of the target to be identified. Specifically, the original scale value of each vertex in the initial features of the target to be identified is multiplied by the scale correction coefficient, thereby stretching or compressing the feature scale of the target to be identified to the same scale order as the reference feature. The formula for calculating the corrected scale is as follows: in, Indicates the target to be identified. Correction scale for each vertex Indicates the target to be identified. The original scale of each vertex.

[0074] Finally, the wavenumbers of each vertex in the initial features of the target to be identified are combined with the corrected scale to generate the corrected features of the target. Since the wavenumber axis corresponds to the physical spectral wavelength position and usually has absolute stability, only the scale dimension is normalized and corrected, while the original wavenumber values ​​are retained. The final corrected features of the target to be identified are a set of coordinate points composed of the original wavenumber and the corrected scale.

[0075] In a specific embodiment, the corrected features of the target to be identified are matched with the current reference features to generate the matching cost of the current reference features. The specific execution process includes: First, a feature vector to be identified is constructed based on the wavenumber and modified scale of each vertex in the modified features of the target to be identified; simultaneously, a current reference feature vector is constructed based on the wavenumber and scale of each vertex in the current reference features. The feature vector is a numerical sequence of discrete feature point coordinates arranged in a preset order (e.g., in ascending order of wavenumber), and the feature vector can characterize the spectral topology of the target in multiple dimensions.

[0076] Subsequently, the Euclidean distance between the feature vector to be identified and the current reference feature vector is calculated to generate the matching cost of the current reference feature. The Euclidean distance intuitively measures the geometric distance between the target and the reference target in the multidimensional feature space; the smaller the distance value, the more similar the spectral structures of the two targets are. The formula for calculating the matching cost is as follows: in, This represents the matching cost of the current reference feature. This indicates the number of feature points involved in the matching. Indicates the first feature vector to be identified Wavenumber of each feature point This represents the th element in the current reference eigenvector. Wavenumber of each feature point Indicates the first feature vector to be identified Correction scale for each feature point This represents the th element in the current reference eigenvector. The scale of each feature point.

[0077] Through the above step S4, the features to be identified are normalized using the correction coefficient based on the maximum scale, and similarity is measured using Euclidean distance. This effectively eliminates the interference of spectral scale deformation caused by environmental factors, enabling spectral data collected under different physical conditions to be accurately compared with reference data in the standard library, significantly improving the accuracy and robustness of target recognition in complex environments.

[0078] Step S5: Select the reference feature with the smallest matching cost from the matching costs of each reference feature as the candidate feature; determine the recognition result of the target to be identified based on the candidate feature and the preset discrimination threshold.

[0079] In a preferred embodiment, the identification result of the target to be identified is determined based on candidate features and a preset discrimination threshold, including: Determine whether the matching cost of the candidate feature is less than the preset discrimination threshold. If so, use the candidate feature as the recognition result of the target to be identified. If not, then the identification result of the target to be identified is determined to be an unknown target.

[0080] Specifically, after generating the matching cost of the target to be identified relative to each reference feature in the preset reference feature database, all calculated matching costs are compared. The matching cost with the smallest value is identified from all matching costs. The matching cost with the smallest value represents the reference target with the smallest difference in spectral feature structure from the target to be identified under the current scale correction and feature matching rules. The reference feature corresponding to the matching cost with the smallest value is determined as a candidate feature. The candidate feature only represents the option most similar to the target to be identified in the current preset reference feature database, but the candidate feature is not necessarily equivalent to the true identity of the target to be identified, because the target to be identified may belong to an unknown category outside the preset reference feature database. Therefore, a preset discrimination threshold needs to be introduced to verify the credibility of the identification result.

[0081] In a preferred embodiment, the identification result of the target to be identified is determined based on candidate features and a preset discrimination threshold. The specific execution process includes: First, it is determined whether the matching cost of the candidate feature is less than a preset discrimination threshold. The preset discrimination threshold is a quantitative standard used to measure the reliability of the recognition result. In specific implementation settings, the preset discrimination threshold is a fixed value pre-set based on empirical data or statistical analysis. To ensure the reasonableness of the threshold setting, the preset discrimination threshold is set to a range of one percent to ten percent of the baseline matching cost. The baseline matching cost refers to the maximum matching cost value obtained by matching a standard test target with all samples in the preset reference feature database. By setting the above percentage range, the preset discrimination threshold can effectively filter out erroneous matches that, although ranking first in similarity in the preset reference feature database, still have significant actual differences.

[0082] If so, meaning the matching cost of the candidate feature is less than the preset discrimination threshold, it indicates that the difference between the target to be identified and the candidate feature is within the allowable error range, and the target to be identified and the candidate feature have extremely high structural consistency. Therefore, the reference target category information corresponding to the candidate feature is output as the identification result of the target to be identified, and the identification is considered successful.

[0083] If not, meaning the matching cost of the candidate feature is greater than or equal to the preset discrimination threshold, it indicates that the difference between the target and the candidate feature exceeds the allowable error range. Even if the candidate feature is the closest item to the target in the preset reference feature database, the similarity between the target and the candidate feature still does not meet the standard for identifying them as the same target. Therefore, the identification result of the target is determined to be an unknown target, that is, the target is determined not to belong to the scope covered by the preset reference feature database.

[0084] Through the above step S5, combined with the dual confirmation mechanism of minimum matching cost optimization and threshold decision, it is possible not only to accurately identify targets from the preset reference feature database, but also to effectively reject interfering targets that do not belong to the preset reference feature database, thus avoiding false alarms caused by forced classification and significantly improving the confidence and reliability of target identification results.

[0085] For example, the construction process of the preset reference feature database and the specific data structures contained in the preset reference feature database are executed as follows: The pre-defined reference feature database is a collection of standard spectral fingerprint features for known target categories. It provides benchmark data support for matching and classifying targets to be identified. The construction process of the pre-defined reference feature database strictly follows the logical steps of standard sample collection, unified feature transformation, and structured storage.

[0086] First, spectral data of multiple known categories of standard targets are acquired. These standard targets encompass all objects of interest that may appear in the pre-defined detection mission. Specifically, they include gas samples with different chemical compositions, camouflage coating samples of different materials, and samples of specific types of ground features. To ensure the standardization of the reference data, the spectral detection of the standard targets is conducted under controlled standard atmospheric pressure and standard room temperature conditions, thereby obtaining high-precision standard reference spectral data. The standard reference spectral data is mathematically more accurate than the spectral data of the targets to be identified acquired in the field.

[0087] To ensure the physical comparability between the reference features and the features of the target to be identified, the same multi-scale spatial transformation and vertex extraction logic used for the target must be employed to process the standard reference spectral data. Specifically, a multi-scale spatial transformation is performed on each standard reference spectral data point, generating a corresponding reference feature map using a pre-defined Gaussian kernel function and scale sequence. Subsequently, contour tracing is performed on the reference feature map to extract all closed arch structures, and the vertex coordinates of each arch structure are further identified. The wavenumber and scale values ​​of all arch structure vertices corresponding to each standard target are combined to form a reference feature set for the standard target. Simultaneously, by filtering the scale values ​​of all vertices in the reference feature set, the scale value of the vertex with the largest scale value is marked as the maximum scale of the reference feature.

[0088] The processed data is stored in a storage medium according to a preset data structure, forming a preset reference feature database. Each storage unit in the preset reference feature database corresponds to a specific standard target. Each storage unit specifically contains target category identification information, a reference feature vector table, and a reference maximum scale parameter. The target category identification information is used to uniquely identify the target category name or code to which the reference feature belongs. The reference feature vector table records a list of coordinates of all arched structure vertices of the standard target, with each coordinate item containing a defined wavenumber position and a defined scale. The reference maximum scale parameter stores the maximum scale value among all feature points of the standard target separately. The reference maximum scale parameter is used as the denominator or numerator reference for calculating the scale correction coefficient in subsequent matching steps.

[0089] To more intuitively illustrate the contents of the preset reference feature database, the following specific numerical storage example is set. Assume the preset reference feature database stores a reference target with the ID REF_001. The target category identifier field stores the information as polyethylene material. The reference feature vector table field records three key feature points corresponding to REF_001: feature point A, feature point B, and feature point C. Feature point A has a wavenumber position of 1450 and a scale of 6.5. Feature point B has a wavenumber position of 2850 and a scale of 12.0. Feature point C has a wavenumber position of 720 and a scale of 3.5. By comparing the scale of feature point A (6.5), feature point B (12.0), and feature point C (3.5), the value 12.0 is determined to be the maximum value. Therefore, the value 12.0 is stored in the reference maximum scale parameter field. When performing target recognition, the stored feature point data and the maximum scale parameter 12.0 will be directly called and used in the process of scale correction and matching cost calculation with the target to be recognized.

[0090] By pre-constructing a preset reference feature database containing precise coordinates and statistical extreme values, the online recognition process can quickly index accurate matching benchmarks, eliminating the need to repeatedly perform feature extraction calculations on the reference data during real-time processing, thus significantly improving the efficiency of target recognition.

[0091] To further illustrate the robustness of the method provided in this application against changes in environmental physical conditions, the following specific application scenario is set: The standard target spectral data stored in the preset reference feature database is collected at standard room temperature, while in actual field detection scenarios, the target to be identified is in a high-temperature environment. High temperatures intensify the molecular thermal motion of the target, causing nonlinear broadening of its spectral absorption or emission features along the wavenumber axis. If a traditional direct matching method is used, the broadened spectrum will produce a significant matching error with the reference spectrum at standard room temperature. However, using the method provided in this embodiment, after multi-scale spatial transformation, the spectral features of the target broadened due to high temperature are represented in the feature map as a broadened arched structure at the bottom, while the topological relationship at the apex of the arched structure remains intact. After extracting the initial features of the target, the maximum scale of the target is identified, and the ratio of the maximum scale of the corresponding reference target in the preset reference feature database to the maximum scale of the target is calculated to generate a scale correction coefficient. The scale correction coefficient can accurately stretch or compress the features of the target that have undergone nonlinear scale deformation due to high temperature back to the standard scale reference. The subsequent Euclidean distance matching calculation mathematically eliminates the nonlinear broadening interference caused by ambient temperature, thus accurately identifying the true category of high-temperature targets. This correction mechanism is also applicable to spectral nonlinear stretching and deformation caused by changes in ambient pressure and detection distance, and has strong engineering practical value.

[0092] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0093] like Figure 2 As shown, an embodiment of the present invention provides a target recognition device based on feature map, including: a data acquisition module, a multi-scale spatial transformation module, a feature correction module, and a recognition and determination module; The data acquisition module is used to acquire the spectral data of the target to be identified; wherein, the spectral data is used to characterize the spectral intensity of the target to be identified at different wavenumbers; The multi-scale spatial transformation module performs multi-scale spatial transformation on the spectral data of the target to be identified, generating a feature map of the target; wherein, the feature map contains several arched structures used to characterize the spectrum as it changes with scale. The feature correction module is used to extract the vertices of each arch structure from the feature map; determine the wavenumber and scale of each arch structure vertex in the feature map as the initial features of the target to be identified; take the maximum scale of each arch structure vertex in the feature map as the maximum scale of the target to be identified; for each reference feature in the preset reference feature library, determine the maximum scale of the current reference feature according to the scale of each arch structure vertex corresponding to the current reference feature; based on the maximum scale of the current reference feature and the maximum scale of the target to be identified, correct the scale of the initial feature of the target to be identified to generate the corrected feature of the target to be identified; perform matching calculation between the corrected feature of the target to be identified and the current reference feature to generate the matching cost of the current reference feature; wherein, the reference feature includes the wavenumber and scale of each arch structure vertex corresponding to the reference target; The identification and determination module is used to select the reference feature with the smallest matching cost from the matching costs of each reference feature as a candidate feature; and to determine the identification result of the target to be identified based on the candidate feature and a preset discrimination threshold.

[0094] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the target recognition method based on feature maps described above in any one of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0095] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0096] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A target recognition method based on feature map, characterized in that, include: Acquire spectral data of the target to be identified; wherein, the spectral data is used to characterize the spectral intensity of the target to be identified at different wavenumbers; A multi-scale spatial transformation is performed on the spectral data of the target to be identified to generate a feature map of the target; wherein, the feature map contains several arched structures used to characterize the spectrum as it changes with scale. The vertices of each arch structure are extracted from the feature map; the wavenumber and scale of each arch structure vertex in the feature map are determined as the initial features of the target to be identified; the maximum scale of each arch structure vertex in the feature map is taken as the maximum scale of the target to be identified. For each reference feature in the preset reference feature library, the maximum scale of the current reference feature is determined according to the scale of the vertices of each arch structure corresponding to the current reference feature; based on the maximum scale of the current reference feature and the maximum scale of the target to be identified, the scale of the initial feature of the target to be identified is corrected to generate the corrected feature of the target to be identified; the corrected feature of the target to be identified is matched with the current reference feature to generate the matching cost of the current reference feature; wherein, the reference feature includes the wave number and scale of the vertices of each arch structure corresponding to the reference target; From the matching costs of each reference feature, the reference feature with the lowest matching cost is selected as the candidate feature; based on the candidate feature and the preset discrimination threshold, the recognition result of the target to be identified is determined.

2. The target recognition method based on feature map as described in claim 1, characterized in that, Multi-scale spatial transformation is performed on the spectral data of the target to be identified to generate a feature map of the target, including: Based on a preset Gaussian kernel function, convolution operations are performed on the spectral data of the target to be identified at preset scales to generate convolution results corresponding to each scale. For each scale, calculate the second derivative of the convolution result with respect to the wavenumber corresponding to the current scale to generate the current second derivative curve; extract the zero-crossing points of the current second derivative curve from the current second derivative curve as the zero-crossing points corresponding to the current scale. A two-dimensional coordinate system is constructed with wavenumber as the abscissa and scale as the ordinate; The zero-crossing points corresponding to each scale are mapped to a two-dimensional coordinate system to generate a feature map of the target to be identified.

3. The target recognition method based on feature map as described in claim 2, characterized in that, The vertices of each arch structure are extracted from the feature map, including: Contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing point contour lines; The outlines of each zero-crossing point are defined as the arched structures; For each arch structure, select the zero-crossing point with the largest ordinate from the zero-crossing points corresponding to the current arch structure, and use it as the vertex of the current arch structure.

4. The target recognition method based on feature map as described in claim 3, characterized in that, Contour tracing is performed on each zero-crossing point in the feature map to identify several closed zero-crossing contour lines, including: Repeat the contour extraction operation until all zero-crossing points in the feature map are marked, generating several closed zero-crossing contour lines; The contour extraction operation includes: Based on the preset traversal order, the feature map is scanned, and the first zero-crossing point that has not been marked as processed is taken as the current starting node. Based on the current starting node, determine the current search path; add the current search path to the search path set, and mark all zero-crossing points on the current search path as processed; Determine if there are any unmarked zero-crossing points in the feature map. If not, perform a closure check on each search path in the search path set to determine the contour lines of each closed zero-crossing point. If so, scan the feature map based on the preset traversal order and update the first unmarked zero-crossing point found to the current starting node.

5. The target recognition method based on feature map as described in claim 4, characterized in that, Based on the current starting node, determine the current search path, including: Add the current starting node to the current connection path; Repeat the adjacent zero-crossing point connection operation until there are no zero-crossing points that have not been added to the current connection path and have not been marked as processed, and generate the current search path; The adjacent zero-crossing point connection operation includes: Within the preset neighborhood of the current connected node, search for any zero-crossing points that have not been added to the current connected path and have not been marked as processed; If such a zero-crossing point exists, it will be selected as a candidate zero-crossing point within the preset neighborhood of the current connected node that has not been added to the current connected path and has not been marked as processed; where the initial connected node is the current starting node. Based on the candidate zero-crossing points, construct a set of neighborhood zero-crossing points; Based on the preset direction search priority, a unique target zero-crossing point is determined from the set of neighboring zero-crossing points; the target zero-crossing point is added to the current connection path, and the target zero-crossing point is updated as the current connection node; If it does not exist, the current connection path will be used as the current search path.

6. The target recognition method based on feature map as described in claim 5, characterized in that, Perform a closure check on each search path in the search path set to determine the zero-crossing contour lines of each closed path, including: For each search path in the set of search paths, determine the start and end nodes of the current search path; Determine whether the starting node of the current search path is within the preset neighborhood of the ending node of the current search path; If so, then determine that the current search path satisfies the closure condition, and define the current search path as a closed zero-crossing contour line; If not, then the current search path does not meet the closure condition; If none of the search paths meet the closure condition, calculate the confidence level of each search path based on the start and end nodes of the search path; sort the search paths in descending order based on the confidence level to generate a path sorting result; select the first preset number of search paths from the path sorting result as each path to be compensated; perform closure connection processing on each path to be compensated to generate a closed zero-crossing contour line.

7. The target recognition method based on feature map as described in claim 6, characterized in that, Based on the maximum scale of the current reference features and the maximum scale of the target to be identified, the scale of the initial features of the target to be identified is corrected to generate corrected features of the target to be identified, including: Calculate the ratio of the maximum scale of the current reference feature to the maximum scale of the target to be identified, and generate the current scale correction coefficient; Based on the current scale correction coefficient, the scale of each vertex in the initial features of the target to be identified is corrected to generate the corrected scale of each vertex in the initial features of the target to be identified. The wavenumbers of each vertex in the initial features of the target to be identified are combined with the correction scale to generate the corrected features of the target to be identified.

8. The target recognition method based on feature map as described in claim 7, characterized in that, The corrected features of the target to be identified are matched with the current reference features to generate the matching cost of the current reference features, including: Based on the wave number and correction scale of each vertex in the modified features of the target to be identified, construct the feature vector to be identified; Construct the current reference feature vector based on the wave number and scale of each vertex in the current reference feature; Calculate the Euclidean distance between the feature vector to be identified and the current reference feature vector, and generate the matching cost of the current reference feature.

9. The target recognition method based on feature map as described in claim 8, characterized in that, Based on candidate features and a preset discrimination threshold, the recognition result of the target to be identified is determined, including: Determine whether the matching cost of the candidate feature is less than the preset discrimination threshold. If so, use the candidate feature as the recognition result of the target to be identified. If not, then the identification result of the target to be identified is determined to be an unknown target.

10. A target recognition device based on feature map, characterized in that, include: The module consists of a data acquisition module, a multi-scale spatial transformation module, a feature correction module, and a recognition and judgment module. The data acquisition module is used to acquire the spectral data of the target to be identified; wherein, the spectral data is used to characterize the spectral intensity of the target to be identified at different wavenumbers; The multi-scale spatial transformation module performs multi-scale spatial transformation on the spectral data of the target to be identified, generating a feature map of the target; wherein, the feature map contains several arched structures used to characterize the spectrum as it changes with scale. The feature correction module is used to extract the vertices of each arch structure from the feature map; determine the wavenumber and scale of each arch structure vertex in the feature map as the initial features of the target to be identified; take the maximum scale of each arch structure vertex in the feature map as the maximum scale of the target to be identified; for each reference feature in the preset reference feature library, determine the maximum scale of the current reference feature according to the scale of each arch structure vertex corresponding to the current reference feature; based on the maximum scale of the current reference feature and the maximum scale of the target to be identified, correct the scale of the initial feature of the target to be identified to generate the corrected feature of the target to be identified; perform matching calculation between the corrected feature of the target to be identified and the current reference feature to generate the matching cost of the current reference feature; wherein, the reference feature includes the wavenumber and scale of each arch structure vertex corresponding to the reference target; The identification and determination module is used to select the reference feature with the smallest matching cost from the matching costs of each reference feature as a candidate feature; and to determine the identification result of the target to be identified based on the candidate feature and a preset discrimination threshold.

Images